Selective noise tolerant mode of operation in memory
Adjusting the memory input power level through selective noise allowable mode, combined with data quantization technology, the problems of volatile memory in terms of power consumption and data stability are solved, achieving more efficient memory operation and longer battery life.
Patent Information
- Application Number
- CN201810213457.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2017-03-30
- Filing Date
- 2018-03-15
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2038-03-15
AI Technical Summary
Existing volatile memory has great challenges in power consumption, especially in mobile applications and stationary systems that consume a lot of power, and data from volatile memory may be lost when power is lost or system resets.
A memory operating system that adopts a selective noise allowable mode reduces power consumption by adjusting the input power level of the memory, and combines data quantization technology to allow a certain degree of noise to improve performance and reduce power consumption.
Effectively reduces power consumption in memory operation, improves performance, and extends battery life in mobile applications while reducing the risk of data loss.
Smart Images

Figure CN108694019B_ABST
Abstract
Description
Technical Field
[0001] Certain embodiments of the present description generally relate to management of memory resources. Background Art
[0002] Non-volatile memories that are capable of storing large amounts of data frequently have disadvantages such as slower input / output speeds compared to volatile, smaller capacity memories. In contrast, smaller, volatile memories that can operate faster may have greater costs per bit basis for storing data than larger capacity non-volatile memories. However, data stored in volatile memories may be lost in the event of a loss of power or a system reset. Therefore, power is typically continuously supplied to the volatile memory as long as the data stored in the volatile memory is to be maintained.
[0003] One type of volatile memory is dynamic random access memory (DRAM), where data stored in the memory is typically "refreshed" in periodic refresh operations. Another type of volatile memory is static random access memory (SRAM), which is able to maintain data without a refresh operation, but nonetheless typically requires a continuous supply of power to maintain the data stored in the SRAM memory. As a result, the power consumption of volatile memory is often a concern, not only in mobile applications with limited power supplies, but also in stationary systems where the memory consumes a large amount of power. BRIEF DESCRIPTION OF THE DRAWINGS
[0004] Embodiments of the present disclosure are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings in which like references refer to similar elements.
[0005] Figure 1 Depicted is a high-level block diagram illustrating one embodiment of a system employing a selective noise tolerant mode of memory operation in accordance with the present Description.
[0006] Figure 2a-2b and Figure 3a-3b Various hierarchical levels of data storage of the memory of FIG. 2 are depicted.
[0007] Figure 4 Embodiments of a memory interface that employs a selective noise tolerant mode of memory operation in accordance with the present description are depicted.
[0008] Figure 5 is a data structure depicting an example of a selective noise tolerant mode of memory operation and associated memory input power levels according to the present Description.
[0009] Figure 6Embodiments of data transfer logic employing a selective noise tolerant mode of memory operation according to the present description are depicted.
[0010] Figure 7 is a data structure depicting an example of a selective noise tolerant mode of memory operation associated with various workload processing tiers according to the present Description.
[0011] Figure 8 Embodiments of a memory employing a selective noise tolerant mode of memory operation with associated memory input power levels in accordance with the present description are depicted.
[0012] Fig. 9 Depicted are examples of dynamically changing memory input power levels in association with dynamically changing a selective noise tolerant mode of memory operation in accordance with the present description.
[0013] Fig.10 An embodiment of a memory employing multiple memory banks with selective noise tolerant modes of memory operation having associated memory input power levels in accordance with the present description is depicted.
[0014] Fig.11 Another embodiment of a memory employing multiple memory banks with a selective noise tolerant mode of memory operation having an associated memory input power level in accordance with the present description is depicted.
[0015] Fig.12 Describes the storage Fig.11 The workload of different memory banks in an embodiment processes words of data.
[0016] Fig.13 is a data structure depicting another example of a selective noise tolerant mode of memory operations associated with various memory banks for various workload processing tiers in accordance with the present Description.
[0017] Fig.14 One embodiment of the operation of a system employing a selective noise tolerant mode of memory operation in accordance with the present description is depicted. DETAILED DESCRIPTION
[0018] In the following description, the same reference numerals are given to the same components, regardless of whether they are shown in different embodiments. In order to illustrate one or more embodiments of the present disclosure in a clear and concise manner, the drawings are not necessarily drawn to scale, and certain features may be shown in a slightly schematic form. Features described or shown with respect to one embodiment may be used in the same manner or in a similar manner in one or more other embodiments, or used in combination with or in place of features of other embodiments.
[0019] A system that employs a selective noise tolerant mode of memory operation according to one aspect of the present specification can reduce the level of memory operation power consumption compared to that achieved by many existing devices. In one embodiment, each noise tolerant mode has an associated input power level to the memory. For example, in one embodiment, the greater the tolerance to noise in the data of the processed workload, the more the memory input power is reduced, and the more the power consumption is ultimately reduced.
[0020] Systems employing selective noise tolerant modes of memory operation according to the present specification can be used in conjunction with a variety of different processing workloads having various tolerance levels for noise in the data being processed. For example, deep learning based algorithms and neural networks are commonly used in several machine learning application areas, including feature map recognition, audio / video recognition, video summarization, etc. These workloads currently run on a variety of hardware platforms ranging from, for example, central processing units (CPUs) to graphics processing units (GPUs).
[0021] In order to reduce power consumption and improve performance, performance per watt consumed is increasingly becoming an important differentiator in the field of machine learning applications. For example, various custom hardware accelerators have been proposed to achieve reduced power consumption, improved performance, or both. Other approaches involve various compression techniques for compressing data processed by the processing workload.
[0022] It should be recognized that some algorithms employed in machine learning can tolerate a certain degree of "noise" in the data being processed. As used herein, the term "noise" refers to errors in the data being processed. The errors may be present in the data as originally acquired. For example, an improperly calibrated sensor or defects in the sensor itself may introduce errors in the raw data collected before being processed by the algorithm.
[0023] One method utilizing noise tolerance utilizes quantization of data to reduce the number of bits representing a particular data point. For example, a data point representing a pixel of data or a segment of an audio waveform may be reduced in size, for example, from a data word having a relatively large number of bits to a smaller word having eight bits to improve performance, or reduce power consumption, or both.
[0024] It should be recognized herein that algorithms based on deep learning and other algorithms with multi-layer processing tend to exhibit different levels of noise immunity or noise tolerance from one processing layer to another. For example, the initial processing layer tends to be more tolerant to the noise in the processed data, compared with, for example, an upper processing layer or a subsequent processing layer that tends to be more abstract or more complex. Therefore, the more abstract or complex the processing layer is, the less the layer tends to allow the noise in the processed data. Therefore, the algorithm based on deep learning that utilizes the quantization of data can have various data quantization modes and data formats, so that the initial layer with a higher tolerance to the noise in the data can have improved data quantization, thereby reducing the size of the data to a greater extent compared to the size of the data for the subsequent layers with a lower noise tolerance. Therefore, variable data precision can be used for different processing layers of the algorithm. It should be recognized that such variable precision may impose a significant burden due to the increased complexity of hardware, software or both.
[0025] In one aspect of the specification, it is recognized herein that noise or errors in the data may also be introduced due to data storage with algorithms having multiple processing layers. For example, a bit error in the data may be caused by a bit cell that is not functioning correctly. In order to reduce or eliminate such bit errors, a minimum input voltage of the memory may be specified by the manufacturer of the memory in an attempt to ensure the correct operation of as many bit cells of the memory as possible.
[0026] One aspect of the selective noise tolerance mode of memory operation according to the present specification includes controlling multiple levels of power supply of the memory in a selectable memory input power mode according to the selected noise tolerance mode of the memory operation. Thus, the memory input power mode can be selected from multiple memory input power modes according to the noise tolerance level of the workload processing layer executed.
[0027] Various embodiments are described herein in terms of selecting an input power level for a memory. It should be appreciated that a variety of different electrical parameters may be used to select the input power, which in turn affect the power consumption level of the memory when changed. For example, selecting an input voltage level for a memory herein is considered to be selecting an input power level when the term is used herein, because the input voltage level affects the power consumption of the memory. As another example, selecting an input current level for a memory herein is considered to be selecting an input power level when the term is used herein, because the input current level affects the power consumption of the memory. It should be appreciated that a variety of different techniques may be used to select the input power level depending on the specific application.
[0028] In another aspect of the present specification, various parameters of the selective noise tolerance mode of memory operation according to the present specification can be programmable. For example, a specific memory input power level can be programmed by setting a register or other data structure to store a value representing one or more programmable input power levels. As another example, noise tolerance levels and memory input power levels associated with various workload processing layers can be programmed in a similar manner.
[0029] Various parameters of the selective noise tolerant mode of memory operation according to the present specification may be programmed when a memory controller or other device employing the selective noise tolerant mode of memory operation is manufactured. Additionally, in some embodiments, system users and applications running on the system may be provided with the ability to program one or more parameters of the selective noise tolerant mode of memory operation according to the present specification (e.g., those described above) in real time, for example, while the system is in use.
[0030] The selective noise tolerant mode of memory operation according to the present specification can be employed alone or in combination with, for example, other techniques (e.g., data quantization with various data quantization modes and data formats). Thus, an initial layer having a higher tolerance for noise in the data can have a selected noise tolerant mode of operation combined with increased data quantization to improve performance, reduce power consumption, or both.
[0031] Implementations of the described techniques may include hardware, methods or processes, or computer software on a computer-accessible medium. Thus, embodiments include corresponding computer systems, apparatuses, and computer programs recorded on one or more computer storage devices, each of which is configured to perform the actions of the method.
[0032] A memory interface that employs a selective noise tolerant mode of memory operation according to the present specification is described herein in conjunction with data sectors, blocks of sectors, regions of blocks, and storage bodies of regions. However, it should be appreciated that the memory interface according to the present specification can be applied to other units and sub-units of data, such as volumes, tracks, segments, files, bytes, etc.
[0033] The memory interface with the selective noise tolerant mode of memory operation according to the present specification can be used in one embodiment for a system of one or more computers configured to perform specific operations or actions of the selective noise tolerant mode of memory operation by means of software, firmware, hardware or a combination thereof installed on the system that causes or causes the system to perform actions when in operation. One or more computer programs can be configured to perform specific operations or actions authorized by the selective memory input power mode by means of instructions that cause the device to perform actions when executed by a data processing device.
[0034] Other embodiments include corresponding computer systems, apparatuses, and computer programs recorded on one or more computer storage devices, each of which is configured to perform the actions of the method. Implementations of the described techniques may include hardware, methods or processes, or computer software on a computer-accessible medium.
[0035] Thus, the operations described herein are performed by logic configured to automatically or substantially automatically perform operations (except where indicated to be performed manually, e.g., user selections) with minimal or no system operator intervention. Thus, as used herein, the term "automatically" includes fully automatic, i.e., operations are performed by one or more hardware or software controlled machines without human intervention (e.g., user input to a graphical user selection interface). As used herein, the term "automatically" also includes primarily automatic, i.e., a majority of operations (e.g., greater than 50%) are performed by one or more hardware or software controlled machines without human intervention (e.g., user input to a graphical user selection interface), and the remainder of the operations (e.g., less than 50%) are performed manually, i.e., manual operations are performed by one or more hardware or software controlled machines with human intervention (e.g., user input to a graphical user selection interface) to guide the execution of the operations.
[0036] Many of the functional elements described in this specification have been labeled "logic" in order to more particularly emphasize their implementation independence. For example, logic elements can be implemented as hardware circuits including custom very large scale integration (VLSI) circuits or gate arrays, off-the-shelf semiconductors (e.g., logic chips, transistors), or other discrete components. Logic elements can also be implemented in programmable hardware devices such as field programmable gate arrays, programmable array logic, programmable logic devices, etc.
[0037] Logical elements can also be implemented in software for execution by various types of processors. A logical element comprising executable code may, for example, comprise one or more physical or logical blocks of computer instructions, which may, for example, be organized as an object, a program, or a function. Nevertheless, the executable files of the identified logical elements need not be physically located together, but may comprise disparate instructions stored in different locations, which, when logically joined together, comprise the logical element and achieve the stated purpose for the logical element.
[0038] In fact, the executable code for the logic element can be a single instruction or many instructions, and can even be distributed over several different code segments, distributed among different programs, distributed among different processors, and distributed across several memory devices. Similarly, operational data can be identified and described herein within the logic element, and can be embodied in any suitable form and organized into any suitable type of data structure. The operational data can be collected as a single data set, or can be distributed over different locations, including over different storage devices.
[0039] Figure 1 1 is a high-level block diagram illustrating selected aspects of a computing system implemented according to an embodiment of the present disclosure. System 10 may represent any of a plurality of electronic or other computing devices, which may include a memory device. Such electronic devices may include cloud storage systems and other computing devices, such as hosts, servers, personal computers, workstations, telephone devices, network devices, virtualization devices, storage controllers, portable or mobile devices (e.g., laptops, netbooks, tablet computers, personal digital assistants (PDAs), portable media players, portable gaming devices, digital cameras, mobile phones, smart phones, feature phones, etc.) or components (e.g., systems on chips, processors, bridges, memory controllers, memories, etc.). System 10 may be powered by a battery, a renewable power source (e.g., a solar panel), wireless charging, or by using an AC outlet.
[0040] In alternative embodiments, system 10 may include more elements, fewer elements, and / or different elements. In addition, although system 10 may be depicted as including separate elements, it should be appreciated that such elements may be integrated onto a platform, such as a system on a chip (SoC). In the illustrated example, system 10 includes a central processing unit or microprocessor 20, a memory controller 30, a memory 40, an offload data transfer engine 44, and peripheral components 50, which may include, for example, a video controller, an input device, an output device, a storage device, a network adapter or interface, a power source (including a battery, a renewable power source (e.g., a photovoltaic panel), wireless charging, or a coupling to an AC outlet), etc.
[0041] The microprocessor 20 includes a cache 25, which may be part of a memory hierarchy that stores instructions and data, and the memory 40 may also be part of a memory hierarchy. The microprocessor 20 also includes logic 27, which may include one or more cores, and may include, for example, a system agent. For example, a core typically includes components of the processor involved in executing instructions, such as an arithmetic logic unit (ALU), a floating point unit (FPU), and / or various levels of cache (e.g., L1 and L2 cache). The system agent may include various controllers, such as serial or parallel data path controllers, various levels of cache (e.g., L3 cache), snoop agent pipelines, on-die memory controllers, and / or other logic.
[0042] Communications between the microprocessor 20 and the memory 40 may be facilitated by a memory controller (or chipset) 30, which may also facilitate communications with peripheral components 50. For example, the memory controller 30 may be on-die with components such as the microprocessor 20, the offload engine 44, memory in the memory 40, or components in the peripheral components 50, or may be separate.
[0043] The peripheral component 50 as a storage device can be, for example, a non-volatile storage device, such as a solid-state drive (SSD), a disk drive including a redundant array of independent disks (RAID), an optical drive, a tape drive, a flash memory, etc. The storage device may include an internal storage device or an attached or network accessible storage device. The microprocessor 20 is configured to write data to the memory 40 and read data from the memory 40. The program in the storage device is loaded into the memory and executed by the processor. The offload data transfer engine 44 facilitates memory-to-memory data transfer, which bypasses the microprocessor to reduce the load of such transfer on the microprocessor 20. As explained in more detail below, an embodiment of a memory interface with a selective noise-tolerant mode of memory operation according to the present specification can provide a suitable power reduction level suitable for a specific application.
[0044] The network controller or adapter enables communication with networks such as Ethernet, fiber channel arbitration loop, etc. In addition, in some embodiments, the architecture may include a video controller configured to display information represented by data in the memory on a display monitor, wherein the video controller may be embodied on a video card or integrated into an integrated circuit assembly mounted on a motherboard or other substrate. The input device is used to provide user input to the processor, and may include a keyboard, a mouse, a stylus, a microphone, a touch-sensitive display screen, an input pin, a socket, or any other activation or input mechanism known in the art. The output device can visualize information sent from the processor or other components such as a display monitor, a printer, a storage device, an output pin, a socket, etc. The network adapter may be embodied on a network card such as a peripheral component interconnect (PCI) card, PCI-Express, or some other I / O card, or on an integrated circuit assembly mounted on a motherboard or other substrate. For example, the peripheral component 50 may also include an RF receiver / transmitter, such as in a mobile phone embodiment. Additional examples of peripheral components 50 that may be provided in the system include audio devices and temperature sensors that transmit temperature updates for storage in memory.
[0045] One or more of the components of system 10 may be omitted depending on the particular application. For example, a network router may lack, for example, a video controller.
[0046] The system 10 includes a cache 25, a memory 40, a memory controller 30, and a peripheral component 50. Figure 1Any one or more of the devices may include a memory using a memory interface having a selective noise tolerant mode of memory operation according to the present specification, including a random access memory such as a DRAM or SRAM memory. The memory using the memory interface may also be embodied as any type of data storage device capable of storing data in a persistent manner (even until power to the memory is interrupted), such as but not limited to a byte-addressable write-in-situ non-volatile memory, a ferroelectric transistor random access memory (FeTRAM), a nanowire-based non-volatile memory, a magnetoresistive random access memory (MRAM), a spin transfer torque (STT)-MRAM, a phase change memory (PCM), a storage class memory (SCM), a universal memory, Ge2Sb2Te5, a programmable metallization cell (PMC), a resistive memory (RRAM), a RESET (amorphous) cell, a SET (crystalline) cell, PCME, an Austenite memory, a ferroelectric memory (also known as a polymer memory and poly (N-vinyl carbazole)), a ferromagnetic Memory (also known as spintronics, SPRAM (spin transfer torque RAM), STRAM (spin tunnel RAM), magnetic memory, magnetic random access memory (MRAM), and semiconductor-oxide-nitride-oxide semiconductor (SONOS, also known as dielectric memory) according to the embodiments described herein. This type of memory according to the embodiments described herein can be used in a stand-alone memory circuit or logic array, or can be embedded in a microprocessor and / or digital signal processor (DSP). In addition, it should be noted that although the systems and processes are described herein primarily with reference to memory types used in microprocessor-based systems in illustrative examples, it should be recognized that certain aspects, architectures, and principles of the present disclosure are equally applicable to other types of memories used in other types of logic devices in view of the disclosure herein.
[0047] One or more of the memory devices of memory 40 and peripheral component 50 may have a rectangular or orthogonal array of rows and columns of cells (e.g., bit cells), where each bit cell is configured to store a bit state. The array of bit cells may be logically subdivided into arrays 52 ( Figure 2a Depending on the size of the memory, the bit cell array may have tens, hundreds, thousands, or more of such regions 54. Regions 54 may be logically subdivided into array 60 of blocks 70 ( Figure 2b ). Depending on the size of the memory, the array of blocks may have tens, hundreds, thousands, or more such blocks 70. In one embodiment, the memory 40 or the storage device of the peripheral component 50 may include a non-volatile memory such as a flash memory, for example, where each block 70 represents the smallest sub-unit of the memory that can be erased at one time.
[0048] Each block 70 may in turn be subdivided into an array of sectors 74 ( Figure 3a Depending on the size of the memory, a block 70 of sectors 74 may have a single sector or tens, hundreds, thousands, or more of such sectors 74. Each sector 74 may in turn be subdivided into an array of memory locations 80 ( Figure 3b ). Depending on the size of the memory, the sectors 74 of memory locations 80 may have tens, hundreds, thousands, or more such memory locations 80. A specific example of a sector is sized enough to store 512 bytes of data. Each memory location includes one or more bit cells that store bits, bytes, words, or other sub-units of data, depending on the particular application. Although the memory interface according to the present specification is described in conjunction with storing data in blocks of one or more sectors, it will be appreciated that other data storage units such as pages, tracks, segments, files, volumes, disks, drives, etc. may be used depending on the particular application.
[0049] Figure 4 1 is a high-level block diagram illustrating selected aspects of another embodiment of a computing system implementing a memory interface 100 according to the present specification. In this embodiment, the computing system includes a plurality of data transfer logic elements DTL1, DTL2 ... DTLn (where n is any positive integer greater than 1), each of which may include a central processing unit, an offload data transfer engine, or other local or remote data transfer logic. For example, each data transfer logic such as data transfer logic DTL1 is configured to generate a memory transaction request (e.g., a read command or a write command) and an address (e.g., a system memory address at which a memory transaction is to occur) and send it to the memory 110 of the interface 100 via an appropriate parallel or serial bus or other data path 102a, 102b ... 102n.
[0050] The memory interface 100 provides an interface between a plurality of data transfer logic elements DTL1, DTL2 ... DTLn and an interface to one or more volatile memories. For example, a volatile memory is a storage medium that requires power to maintain the state of the data stored by the medium. In contrast, a non-volatile memory is a storage medium that does not require power to maintain the state of the data stored by the medium. However, both volatile and non-volatile memories often require input power to execute memory transaction requests such as read commands or write commands to read data from and write data to the memory, respectively.
[0051] In the illustrated embodiment, memory 110 may represent an SRAM-type volatile memory. However, it should be appreciated that the selective noise-tolerant mode of memory operation according to the present specification may be used with various host devices, storage devices, and other memory devices, such as solid-state memory devices using chalcogenide phase change materials (e.g., chalcogenide glass), three-dimensional (3D) cross-point memory, or memory containing memristor technology. It should be appreciated that other types of memory may benefit from a memory interface having a selective noise-tolerant mode of memory operation according to the present specification, depending on the particular application.
[0052] The memory controller 30 of the interface 100 is configured to control input operations and output operations to and from the memory 110, which includes an array of bit cells. Therefore, the memory controller 30 is configured to perform a memory transaction, such as a read operation or a write operation, at a memory address of the memory 110 according to the received system address in response to a memory transaction request and a system address from the data transfer logic elements DTL1, DTL2 ... DTLn.
[0053] The memory controller 30 is coupled to the memory 110 via an appropriate parallel or serial bus or other data path 118. For example, the data path 118 may be a bus where memory transaction requests, memory addresses, error messages, and other data are sent according to an appropriate transaction protocol. The path 118 may also include a power output signal from a multi-stage power supply 174. It should be appreciated that the hardware aspects and transmission protocols of the data paths 102a, 102b ... 102n, 118 may vary depending on the particular application.
[0054] The memory controller 30 includes a selective noise tolerant memory control logic 120 configured to provide a selective memory mode of memory operation according to the present specification. Each memory mode causes a memory, such as the memory 110, to operate according to a noise tolerance level for data of a workload processed by a data transfer logic element, which issues a memory transaction request to the memory for the purpose of reading data in the memory and storing data in the memory. As the noise tolerance level for data of the workload processed by the data transfer logic element changes, the memory mode of the memory 110 can change accordingly. Therefore, in one embodiment, the selective noise tolerant memory control logic 120 includes a noise tolerance mode selection logic 122, which is configured to select a specific noise tolerance mode of the memory according to a specific noise tolerance level for data of the workload processed by the data transfer logic element, which issues a memory transaction request to the memory for the purpose of reading data in the memory or storing data in the memory.
[0055] In one embodiment, the selective noise tolerant memory control logic 120 ( Figure 4 ) also includes command logic 134, which is configured to receive a command from a processor, such as one of the data transfer logics DTL1, DTL2, ... DTLn. Therefore, the noise tolerant mode selection logic 122 is also configured to select a memory mode in response to the command received by the command logic 134. In one embodiment, the command received by the command logic 134 may include a parameter identifying the selected memory mode. Figure 5 is a data structure represented as a graph depicting examples of various memory modes according to the present specification, including, for example, a high noise tolerant memory mode, a medium noise tolerant memory mode, and a low noise tolerant memory mode. Figure 5 The diagram of depicts three examples of memory modes, but it should be appreciated that a system employing a selective noise tolerant mode of memory operation according to the present description may employ fewer or more memory modes depending on the particular application.
[0056] Thus, noise tolerant mode selection logic 122 may be configured to select a noise tolerant memory mode as identified by a memory mode parameter accompanying a command received by command logic 134. In some embodiments, the memory mode parameter identifying the selected noise tolerant mode may be stored in a suitable data structure 138 of command logic 134.
[0057] Figure 6 One embodiment of a processor such as data transfer logic DTL1 is shown, for example, including command logic 150, which is configured to send data to a processor such as memory controller 30 ( Figure 4 ) sends a command including a memory mode parameter identifying a selected noise tolerant mode. The command logic 134 of the receiving memory controller 30 decodes the command, and the noise tolerant mode selection logic 122 of the receiving memory controller 30 selects the noise tolerant mode of the memory controlled by the memory controller 30 according to the memory mode parameter identifying the selected noise tolerant mode in the received and decoded command.
[0058] In this embodiment, the data transfer logic DTL1 also includes a memory transaction logic 156 configured to issue read transactions and write transactions to the memory. As explained in more detail below, the memory of the memory controller that receives the memory mode identification command operates in a selected noise tolerance mode to read, write or store data of the processed workload in the selected noise tolerance mode according to the noise tolerance level of the data of the processed workload.
[0059] In this embodiment, the noise tolerance selection logic 158 ( Figure 6 ) to select a noise tolerance mode of a memory storing data, the noise tolerance selection logic 158 is configured to select the noise tolerance mode according to, for example, the noise tolerance level of data for a workload processed by a data transfer logic element such as the data transfer logic DTL1. The noise tolerance level of data for the workload being processed can be detected by the noise tolerance level detection logic 160. Therefore, the noise tolerance mode selection logic 158 selects the noise tolerance mode according to the noise tolerance level of data for a workload processed by, for example, a data transfer logic element such as the data transfer logic DTL1 detected by the noise tolerance level detection logic 160.
[0060] For example, deep learning based algorithms are often used in machine learning fields such as map recognition, audio / video recognition, video summarization, etc. The processing of these algorithms is often performed in layers where the output of one layer provides input to a higher layer. Figure 7 An example of a processing algorithm with three processing layers is shown: workload processing layer 1, workload processing layer 2, and workload processing layer 3. Algorithms can often tolerate a certain degree of "noise" in the data being processed. As previously mentioned, as the term "noise" is used herein, "noise" refers to errors in the data being processed.
[0061] It should be recognized herein that deep learning based algorithms and other algorithms with multiple layers of processing tend to exhibit different levels of noise immunity or noise tolerance from layer to layer. For example, the initial processing layers (e.g., workload processing layer 1 ( Figure 7 )) For example, they tend to be more tolerant of noise in the data being processed. Thus, the more abstract or complex the processing layer, the less tolerant of noise in the data being processed by that particular layer.
[0062] For example, an algorithm may have an initial layer (e.g., workload processing layer 1) that processes map data representing, for example, a coastline of a landmass. By interpolating two relatively close data points, an expected range of values for the middle data point may be calculated. If the middle data point is outside the expected range, the middle data point may be discarded as "noise" to improve the results of the initial processing layer. In this way, the initial layer has a relatively high tolerance for noise. In contrast, a higher processing level that processes, for example, landmass region data may have a lower tolerance for noise in the processed data.
[0063] Therefore, in one example, the noise tolerance level detection logic 160 ( Figure 6) can detect the noise tolerance level of the workload being processed by identifying a specific layer of the algorithm being processed as the workload of the processor. For example, if the layer being processed is detected by the noise tolerance level detection logic 160 ( Figure 6 ) is detected as an initial layer (e.g., workload processing layer 1), then the noise tolerance level associated with the layer (e.g., high noise tolerance) can be identified, for example, by a programmable lookup table or other data structure of data structure 162, such as by Figure 7 In response, the noise tolerant mode selection logic 158 selects a noise tolerant mode (e.g., a high noise tolerant memory mode ( Figure 5 )), and the command logic 150 issues a command to the memory controller 30 to control the memory storing the data processed by the detected layer to switch the noise tolerant mode of the memory to the selected noise tolerant mode when appropriate.
[0064] In the illustrated embodiment, a lookup table or other programmable data structure is described for associating various processing layers of the algorithm with associated noise tolerance levels. It should be appreciated that other techniques may be employed to identify the noise tolerance levels associated with the processing layers executed. It should also be appreciated that other techniques may be employed to select a specific noise tolerance mode of a memory based on a specific noise tolerance level for data of a workload processed by a data transfer logic element that issues memory transaction requests to the memory for the purpose of reading data in the memory and storing data in the memory.
[0065] exist Figure 6 In the embodiment of the present invention, the command logic 150, the noise tolerance mode selection logic 158, and the noise tolerance level detection logic 160 are implemented, for example, using suitable software, hardware, or a combination thereof associated with a processor such as the data transfer logic DTL1. It should be appreciated that one or more of the software, hardware, or other components of one or more of the command logic 150, the noise tolerance mode selection logic 158, and the noise tolerance level detection logic 160 may be located in whole or in part in other elements of a computer that employs a selective noise tolerance mode of operation according to the present specification. For example, one or more of the software, hardware, or other components of one or more of the command logic 150, the noise tolerance mode selection logic 158, and the noise tolerance level detection logic 160 may be located in whole or in part in the memory controller 30 and its associated software (e.g., a driver of the memory controller 30).
[0066] In another aspect of the specification, the input voltage of the memory is dynamically scaled based on the current workload noise immunity or tolerance (i.e., the noise tolerance or immunity to the current layer of data executed by the processor storing data in the memory and reading data from the memory). In one embodiment, this dynamic scaling of the input power of the memory can dynamically reduce the power consumed by the memory and reduce active leakage.
[0067] exist Figure 4 In the embodiment of the present invention, the selective noise tolerant memory control logic 120 of the memory controller 30 includes a memory input power mode selection logic 170, which is configured to select each input power level mode having an associated programmable input power level ( Figure 5 ) input power level mode. As noted above, the noise tolerance level detection logic 160 ( Figure 6 ) detects the noise tolerance level of the workload layer being processed. In response, the noise tolerance mode selection logic 158 selects a noise tolerance mode compatible with the identified noise tolerance level, and the command logic 150 issues a command to the memory controller 30 to control the memory storing the data processed by the detected layer to switch the noise tolerance mode of the memory to the selected noise tolerance mode when appropriate.
[0068] Therefore, the noise tolerance mode selection logic 122 ( Figure 4 ) selects a memory mode according to the detected noise tolerance level of the workload layer being processed in response to a command received by command logic 134. In response to the memory mode selected by noise tolerance mode selection logic 122, memory input power mode selection logic 170 selects an input power mode having an associated memory input power level. Figure 8 As shown, the multi-stage power supply 174 includes, for example, a regulator 184 (e.g., a low dropout (LDO) regulator), which can be a direct current (DC) linear voltage regulator. The LDO regulator receives power at the power rail input and supplies power to the multi-bank memory 110 ( ) at a selectable, programmable power level according to a control signal ("control") provided by the memory input power mode selection logic 170. Figure 4 ) of the memory bank (in this example, bank 0) output power ( Figure 8 In this way, a computer system according to one aspect of the present specification has logic configured to select an input power level of a memory according to the noise tolerance level of a workload being processed.
[0069] As mentioned earlier, Figure 5is a data structure represented as a graph depicting examples of various noise tolerant memory modes according to the present specification, including, for example, a high noise tolerant memory mode, a medium noise tolerant memory mode, and a low noise tolerant memory mode. Figure 5 As shown in the example of , each noise tolerant memory mode has an associated input power level. In this example, the greater the noise tolerance of the data of the workload, the more the input power of the memory can be reduced to reduce power consumption. It should be recognized that in other embodiments, the relationship between noise tolerance and input power of the memory can vary depending on the specific application.
[0070] Fig. 9 Depicts, for example, a memory (e.g., Figure 8 An example of dynamically changing the input power level of a memory bank 0) as the noise tolerance of the data stored in the memory bank 0 for the workload being processed changes. Therefore, in this example, if the input power level of the data stored in the memory bank 0 for the workload being processed (e.g., Figure 7 If the noise tolerance of the data stored in the memory bank 0 of the workload processing layer 1 is initially at a high tolerance level, then the following may be selected: Figure 5 and Fig. 9 The high noise tolerant memory mode is shown with a low input power level associated with memory bank 0. If the processing layer is changed so that the noise tolerance of the data stored in memory bank 0 for the workload being processed changes from a high tolerance level to a medium tolerance level (e.g., Figure 7 workload processing layer 2 medium noise tolerance level), then you can choose Figure 5 and Fig. 9 The shown has an associated memory bank 0 ( Figure 8 ) is a medium noise tolerant memory mode for medium input power levels. Therefore, bank 0 ( Figure 8 ) can be dynamically raised from the low input power level of the high noise tolerant memory mode to the medium input power level of the medium noise tolerant memory mode. Similarly, if the processing layer is changed again so that the storage of the workload being processed is in the memory bank 0 ( Figure 8 ) changes from a medium tolerance level to a low tolerance level (e.g., Figure 7 workload processing layer 3 low noise tolerance level), you can choose Figure 5 and Fig. 9 The low noise tolerant memory mode shown has a high input power level associated with memory bank 0. Therefore, memory bank 0 ( Figure 8 ) can be dynamically increased from the medium input power level of the medium noise tolerant memory mode to the high input power level of the low noise tolerant memory mode.
[0071] In this way, the input voltage of a memory bank or array of memory banks can be dynamically fine-tuned based on the noise tolerance level for the data of the current workload being processed. As a result, power consumption and active leakage can be reduced. In addition, in mobile applications, battery life can be improved.
[0072] It should be recognized that the reduction in memory input power may introduce some bit errors into the data of the current workload. However, in many applications, it is not expected that these errors will significantly affect the program results, especially in applications with high or relatively high tolerance for noise in the data. For example, the input data representing the feature map typically has a strong spatial correlation level, which gives itself a relatively high noise tolerance level.
[0073] As the overall algorithm being processed advances to subsequent layers that may have a lower tolerance for noise in the data or may otherwise require higher memory input power levels, the memory input power can be adjusted up to the levels required by these workloads. It is believed that depending on the application, a 10% reduction in memory input power level can save on the order of a 20% reduction in dynamic power and leakage at the expense of tolerable memory errors. Furthermore, depending on the specific application, the selective memory input power techniques according to the present specification can be transparent to the algorithm processing software and need not rely on data format conversion.
[0074] For the purpose of providing examples, Fig. 9 A particular sequence of memory modes and associated memory input power levels is depicted. However, it should be appreciated that changes in noise tolerant memory modes and associated memory input power levels may occur dynamically in a variety of different sequences, depending on the particular application. Thus, the memory input power level may dynamically change from a high level to a low level, and vice versa, for example, directly or through a mid-power level.
[0075] Fig.10 A multi-bank embodiment of a memory of a computer system employing selective memory input power according to one aspect of the present specification is depicted. In this embodiment, memory 110 ( Figure 4 ) includes three banks of memory: bank 1, bank 2, and bank 3, each bank storing data of a workload tier having an associated tolerance level for noise of the data stored in the bank. Therefore, the memory input power level provided to each bank (bank 1, bank 2, and bank 3) is a function of the associated tolerance level for noise of the data stored in the bank for the associated workload tier.
[0076] Thus, in this example, data read from and stored in bank 1 is represented by, for example, Figure 7 Therefore, the noise tolerance of the data of workload processing layer 1 stored in storage bank 0 is determined to be at a high noise tolerance level, such as Figure 7 In this example, the input data of workload processing layer 1 is pixel data, as shown by Fig.10 1. The pixel data stored in memory bank 1 has a strong spatial correlation, which lends itself to a relatively high noise tolerance level. Thus, the input pixel data may represent, for example, map features, or video data such as a video file. Therefore, a high noise tolerant memory mode may be selected for memory bank 1, wherein the selected memory mode has an associated low input power level (indicated as "low input power level") for bank 1, as shown in FIG. Figure 5 and Fig.10 shown.
[0077] Input data stored in memory bank 1 is processed by workload processing layer 1, such as Fig.10 and output as output data and stored in the memory bank 2, represented by the arrow marked as "high noise tolerance workload processing layer 1" in the Fig.10 The block representation of pixel data indicated as "Pixel Block 2" in FIG. 1 is a representation of a block of pixel data indicated as "Pixel Block 2" in FIG. The output pixel data processed by workload processing layer 1 and stored in memory bank 2 has a somewhat weak spatial correlation, which lends itself to a medium noise tolerance level. Therefore, a medium noise tolerant memory mode may be selected for memory bank 1, wherein the selected memory mode has an associated medium input power level (indicated as "Medium Input Power Level") for memory bank 2, as shown in FIG. Figure 5 and Fig.10 shown.
[0078] Data stored in memory bank 2, as represented by pixel block 2, is processed by workload processing layer 2, as shown in FIG. Fig.10 The arrows in the figure are indicated by the arrows marked as "medium noise tolerance workload processing layer 2" and are output as output data and stored in the memory bank 2, by Fig.10 3. As with the pixel data of pixel block 2, the pixel data of pixel block 3 has a somewhat weaker spatial correlation than the pixel data of pixel block 1 stored in memory bank 1, and is therefore likewise endowed with its own medium noise tolerance level. Therefore, the previously selected medium noise tolerant memory mode for memory bank 2 is appropriate. As noted above, the selected medium noise tolerant memory mode has an associated medium input power level (indicated as "Medium Input Power Level") for memory bank 2, as shown in FIG. Figure 5 and Fig.10 shown.
[0079] Data stored in memory bank 2 as represented by pixel block 4 for workload processing layer 3 is processed by workload processing layer 3 as shown in FIG. Fig.10 The arrows in the figure are indicated by the arrows marked as "low noise tolerance workload processing layer 3" and are output as output data and stored in the memory storage body 3, by Fig.10 The block representation of pixel data indicated as "Pixel Block 5" in FIG. 1 is a block representation of pixel data indicated as "Pixel Block 5" in FIG. The output pixel data processed by workload processing layer 3 and stored in memory bank 3 has weaker spatial correlation than the pixel data of pixel blocks 2, 3 and 4 stored in memory bank 2, and is therefore also endowed with its own low noise tolerance level. Therefore, a low noise tolerant memory mode for memory bank 3 may be selected, which has an associated high input power level (indicated as "High Input Power Level") for memory bank 3, as shown in FIG. Figure 5 and Fig.10 shown.
[0080] In one embodiment, the noise tolerance level for data stored in the memory banks (bank 1, bank 2, and bank 3) can be relatively constant, so that the memory input power levels (low input power level, medium input power level, and high input power level) respectively provided to the memory banks (bank 1, bank 2, and bank 3) can also remain relatively constant. In other embodiments, if the noise tolerance level of one or more memory banks changes, the memory input power level of the memory bank can be dynamically changed according to the changing noise tolerance level for data stored in the specific memory bank.
[0081] For the purpose of providing examples, Fig.10 A particular arrangement of memory modes and associated memory input power levels for an array of memory banks is depicted. However, it should be appreciated that noise tolerant memory modes and associated memory input power levels may be arranged in a variety of different levels and sequences, depending on the particular application.
[0082] Fig.11 Another multi-bank embodiment of a memory of a computer system employing selective memory input power according to one aspect of the present specification is depicted. In this embodiment, memory 110 ( Figure 4 ) includes two banks of memory: bank 3A and bank 3B, each bank storing a portion of a data word of a workload layer having an associated tolerance level for noise in the data stored in the bank. Fig.12An example of a data word having a first portion (denoted as subunit 1) in which the first portion includes the most significant bits of the word and a second portion (denoted as subunit 0) in which the second portion includes the least significant bits of the word is shown. Subunit 1 of the most significant bits of the word is stored in memory bank 3A, and subunit 0 of the least significant bits of the word is stored in memory bank 3B.
[0083] The memory input power level provided to each bank (bank 3A, bank 3B) is the power level for the workload layers (e.g., Fig.10 The workload of the processing layer 3) is a function of the associated acceptable level of noise in the data. Thus, in one embodiment, Fig.10 The memory bank 3 can be Fig.11 In the embodiment of the present invention, the memory bank 3A is implemented as two memory banks, and the memory bank 3A is used for the workload processing layer 3 having an associated low tolerance level for noise in the data stored in the memory. In addition, the memory input power level provided to each memory bank (memory bank 3A, memory bank 3B) is also a function of the significance of the bits of the word portion stored in the memory bank (memory bank 3A and memory bank 3B), respectively.
[0084] Therefore, in this example, the data read from and stored in the memory bank 3A and the memory bank 3B are represented by, for example, Figure 7 , Fig.10 Therefore, the noise tolerance of the data stored in the memory for the workload processing layer 3 being processed is determined to be such that Figure 7 Thus, a low noise tolerant memory mode of memory bank 3A may be selected, which has an associated high input power level for memory bank 3A (indicated as “high input power level” from voltage regulator LDO3A), as shown in FIG. Fig.11 and Fig.13 As shown, because memory bank 3A contains the most significant bits of the data words stored for workload handling layer 3. Because memory bank 3B stores the least significant bits of the data words stored for workload handling layer 3, a more noise tolerant memory mode may be selected for memory bank 3B, as indicated by the medium input power level (“medium input power level” from voltage regulator LDO3B) input to memory bank 3B, as shown in FIG. Fig.11 and Fig.13As shown. Therefore, it will be appreciated that the reduction in memory input power to memory bank 3B compared to the memory input power level to memory bank 3A may introduce some bit errors into the data of the current workload stored in memory bank 3B. However, in many applications, it is not expected that these errors will significantly affect program results, especially in applications that have a high or relatively high tolerance for noise in the least significant bits of data.
[0085] In a similar manner, storage for Fig.10 The memory bank 2 of the workload processing layer 2 data can be split into two or more banks, which are used to be stored for Fig.10 The workload processes the most significant bit and the least significant bit of the data word of layer 2. If expressed as Fig.13 As indicated in the data structure of the diagram, the memory bank storing the most significant bits of the data word may be operated in the noise tolerant memory mode, as described above in conjunction with Fig.10 As described above, the memory banks storing the least significant bits of the data words for the workload processing layer 2 may be operated in a memory mode having a greater noise tolerance, such as a high noise tolerance, such as Fig.13 as indicated in the diagram.
[0086] exist Fig.13 In the example of , the memory portions storing the most significant bits and the least significant bits of the words of workload processing layer 1 have a selected common noise tolerance level, i.e., for example, a high noise tolerance level. In such an example, depending on the specific application, a single memory bank for storing data words can be utilized.
[0087] exist Fig.11 In the embodiment of the present invention, the multi-stage power supply 174 includes a plurality of regulators, such as represented by LDO regulators (LDO3A, LDO3B), in response to the memory input power mode selection logic 170 ( Figure 4 ) to provide multiple selectable input power levels to various memory banks by appropriate control signals. In this way, a computer system according to one aspect of the present specification has logic configured to select multiple input power levels for one or more memory devices based on the noise tolerance level or multiple noise tolerance levels of the workload or workloads being processed.
[0088] Fig.14 One embodiment of the operation of a computer system employing selective memory input power as a function of noise tolerance level according to the present specification is depicted. In one operation, an initial memory input power level is selected (block 208). Thus, as described above in conjunction with Fig.10As shown, for example, a low input power level, a medium power level, and a high input power level may be selected for memory bank 1, memory bank 2, and memory bank 3, respectively. Such initial memory input power levels may be selected as, for example, default values, or may be selected based on an initially detected noise tolerance level for data stored in the memory banks in conjunction with the workload processing layer executed or to be executed.
[0089] For example, as mentioned above, Figure 4 1, the memory input power mode selection logic 170 selects an input power level mode in which each input power level mode has an associated input power level provided by the multi-stage power supply 174, such as Figure 5 In one example, the noise tolerance level detection logic 160 ( Figure 6 ) detects the noise tolerance level of the workload layer being processed. In response, the noise tolerance mode selection logic 158 selects a noise tolerance mode compatible with the identified noise tolerance level, and the command logic 150 issues a command to the memory controller 30 to control the memory storing the data processed by the detected layer to switch the noise tolerance mode of the memory to the selected noise tolerance mode when appropriate.
[0090] Therefore, the noise tolerance mode selection logic 122 ( Figure 4 ) selects a memory mode based on the detected noise tolerance level of the workload layer being processed in response to a command received by command logic 134. In response to the memory mode selected by noise tolerance mode selection logic 122, memory input power mode selection logic 170 selects an input power mode having an associated memory input power level. The multi-stage power supply, for example, supplies power (at selectable power levels) at selectable power levels based on a control signal (“control”) provided by memory input power mode selection logic 170. Figure 8 The memory bank 1 (denoted as "selectable memory input power level" in the embodiment of the present specification) is output to a memory bank such as memory bank 1. In this way, a computer system according to one aspect of the present specification can select an input power level for the memory based on the noise tolerance level of the workload being processed.
[0091] like Figure 5 As shown in the example of , each noise-tolerant memory mode has an associated input power level. In this example, the greater the noise tolerance of the data of the workload, the more the input power of the memory can be reduced to reduce power consumption.
[0092] A determination is made as to whether the noise tolerance level for the data of the workload processing layer has changed ( Fig.14, block 224). If so, then an appropriate memory input power level is selected (block 230).
[0093] As mentioned earlier, Fig. 9 Describes such Fig.10 An example of dynamically changing the input power level of the memory of memory bank 1 as the noise tolerance of the data stored in memory bank 1 for the workload being processed changes. Therefore, in this example, if the noise tolerance of the data stored in memory bank 1 for the workload being processed (e.g., Figure 7 If the noise tolerance of the data stored in the memory bank 1 of the workload processing layer 1 is initially at a high tolerance level, a high noise tolerance memory mode with a low input power level associated with the memory bank 1 may be selected, such as Figure 5 , Fig. 9 as well as Fig.10 If the noise tolerance of the data stored in the memory bank 1 for the workload being processed changes from a high tolerance level to a medium tolerance level (e.g., Figure 7 The workload of processing layer 2 has a medium noise tolerance level), then the associated memory bank 1 ( Fig.10 ) for a medium noise tolerant memory mode with a medium input power level, such as Figure 5 and Fig. 9 Therefore, memory bank 1 ( Fig.10 ) can be dynamically increased from the low input power level of the high noise tolerant memory mode to the medium input power level of the medium noise tolerant memory mode. Similarly, if the storage of the workload being processed is in memory bank 1 ( Fig.10 ) changes from a medium tolerance level to a low tolerance level (e.g., Figure 7 If the workload of the processing layer 3 has a low noise tolerance level), a low noise tolerance memory mode with a high input power level of the associated memory bank 1 may be selected, such as Figure 5 and Fig. 9 Therefore, memory bank 1 ( Fig.10 ) can be dynamically increased from the medium input power level of the medium noise tolerant memory mode to the high input power level of the low noise tolerant memory mode.
[0094] In this way, the input voltage of a memory bank or array of memory banks can be dynamically fine-tuned based on the noise tolerance level for the data of the current workload being processed. As a result, power consumption and active leakage can be reduced. In addition, in mobile applications, battery life can be improved.
[0095] It should be recognized that the reduction in memory input power may introduce some bit errors into the data of the current workload. However, as explained above, in many applications, it is not expected that these errors will significantly affect the program results, especially in applications with high or relatively high tolerance for noise in the data. For example, the input data representing the feature map typically has a strong spatial correlation, which gives it a relatively high noise tolerance level.
[0096] The foregoing demonstrates that as the overall algorithm being processed advances to subsequent layers that may have a lower tolerance for noise in the data or may otherwise require higher memory input power levels, the memory input power can be adjusted up to the level required by these workloads. Furthermore, depending on the specific application, the selective memory input power techniques according to the present specification can be transparent to the algorithm processing software and need not rely on data format conversion.
[0097] It should be appreciated that the operation depicted in the figure can be performed by a memory interface logic with a selective noise tolerance mode of memory operation utilizing an architecture other than that depicted in the figure and using other types of logic components. The logic components including the logic elements depicted in the figure discussed herein can be configured to perform the described operations using appropriate hardware, software or firmware or various combinations thereof. For example, software can be in the form of firmware, programs, drivers and other instruction sets, and hardware can be in the form of a general logic device such as a microprocessor or a dedicated logic device such as a memory controller, a DMA controller or an engine or an ASIC device.
[0098] The hardware, software or firmware for the device for adopting the selective noise tolerant mode of memory operation according to the present specification can be physically or logically located in any component of the system including the memory itself, a controller such as a memory controller, a DMA controller, a microprocessor, etc. Thus, in one embodiment, one or more of the memory interface logic elements depicted in the figure can be implemented using one or more of the hardware of the memory controller, the firmware of the memory controller, and the software (e.g., the associated driver software of the memory controller). In another embodiment, one or more of the memory interface logic elements depicted in the figure can be implemented, for example, using one or more of the controller hardware such as a central processing unit, or other controllers, firmware for the controller hardware, and software for the controller hardware (e.g., programs and drivers executed by the controller hardware such as the central processing unit). In another embodiment, one or more of the memory interface logic elements depicted in the figure can be implemented, for example, using hardware, firmware or software for offloading both the data transfer engine and the central processing unit.
[0099] It will be appreciated that a device employing a selective noise tolerant mode of memory operation according to the present description may be able to reduce power consumption depending on the particular application.Other aspects may be implemented depending on the particular application.
[0100] Example
[0101] The following examples relate to further embodiments.
[0102] Example 1 is an apparatus for use with a memory, comprising:
[0103] A memory controller having an input power mode selection logic configured to select an input power level of a memory based on a noise tolerance level of a workload, wherein the workload includes a first workload having a first noise tolerance level and a second workload having a second noise tolerance level, and the memory controller also has a multi-stage power supply configured to provide the selected power level to the memory.
[0104] In Example 2, the subject matter of Examples 1-9 (excluding the present example) may optionally include, wherein a second noise tolerance level for a second workload is lower than a first noise tolerance level for a first workload, and wherein a power level input to a memory associated with the second workload is higher than a power level input to a memory associated with the first workload.
[0105] In Example 3, the subject matter of Examples 1-9 (excluding the present example) may optionally include, wherein the input power mode selection logic is configured to dynamically increase the power input to the memory based on a change in the workload from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0106] In Example 4, the subject matter of Examples 1-9 (excluding the present example) can optionally include, wherein the first workload is a first layer of the deep learning neural network and the second workload is a second, subsequent layer of the deep learning neural network.
[0107] In Example 5, the subject matter of Examples 1-9 (excluding the present example) can optionally include, wherein the first layer of the deep learning neural network is configured to process data having a first spatial correlation level, and the second layer of the deep learning neural network is configured to process data having a second spatial correlation level, the second spatial correlation level being weaker than the spatial correlation level of the data having the first spatial correlation level.
[0108] In Example 6, the subject matter of Examples 1-9 (excluding the present example) may optionally include, wherein the memory includes a first memory bank and a second memory bank, and wherein the memory controller is configured to store a data unit having a first sub-unit and a second sub-unit of data bits in the memory, including storing the first sub-unit of bits in the first memory bank of the memory, and storing the second sub-unit of bits in the second memory bank of the memory, wherein the bits of the first sub-unit are more efficient than the bits of the second sub-unit.
[0109] In Example 7, the subject matter of Examples 1-9 (excluding the present example) may optionally include, wherein a multi-stage power supply of the memory controller is configured to provide input power to a first memory bank and a second memory bank of the memory at a selected level, wherein the input power mode selection logic is configured to select the input power level of the first memory bank of the memory to be at a power level that is higher than the selected input power level of the second memory bank of the memory.
[0110] In Example 8, the subject matter of Examples 1-9 (excluding the present example) may optionally include, wherein the memory controller is further configured to dynamically increase the power of at least one of the first memory body and the second memory body input to the memory based on a change in the workload from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0111] In Example 9, the subject matter of Examples 1-9 (excluding this example) may optionally include a system, the system comprising a processor, a memory, the memory controller, and at least one of the following:
[0112] A display is communicatively coupled to the processor, a network interface is communicatively coupled to the central processing unit, and a battery is coupled to provide power to the system.
[0113] Example 10 is a method comprising:
[0114] inputting power to the memory at a first power level,
[0115] Using a processor and a memory, processing a first workload having a first noise tolerance level,
[0116] Using the processor and the memory, processing a second workload having a second noise tolerance level different from the first noise tolerance level, and
[0117] Power is input to the memory at a second power level different from the first power level based on a noise tolerance level of a workload being processed.
[0118] In Example 11, the subject matter of Examples 10-17 (excluding the present example) may optionally include, wherein a second noise tolerance level for the second workload is lower than a first noise tolerance level for the first workload, and wherein a second power level of power input to the memory when the second workload is processed is higher than a first power level of power input to the memory when the first workload is processed.
[0119] In Example 12, the subject matter of Examples 10-17 (excluding the present example) may optionally include dynamically increasing the power input to the memory from a first power level to a second level based on changing the workload processed by the processor and the memory from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0120] In Example 13, the subject matter of Examples 10-17 (excluding the present example) can optionally include, wherein the first workload is a first layer of the deep learning neural network and the second workload is a second, subsequent layer of the deep learning neural network.
[0121] In Example 14, the subject matter of Examples 10-17 (excluding the present example) can optionally include, wherein a first layer of the deep learning neural network processes data having a first spatial correlation level, and a second layer of the deep learning neural network processes data having a second spatial correlation level, the second spatial correlation level being weaker than the spatial correlation level of the data at the first spatial correlation level.
[0122] In Example 15, the subject matter of Examples 10-17 (excluding the present example) may optionally include storing data units in a memory, wherein the storing includes storing a first subunit of bits in a first storage body of the memory, and storing a second subunit of bits in a second storage body of the memory, wherein the bits of the first subunit are more efficient than the bits of the second subunit.
[0123] In Example 16, the subject matter of Examples 10-17 (excluding the present example) may optionally include inputting power to a first memory bank and a second memory bank of the memory, wherein power is input to the first memory bank of the memory at a higher power level than power input to the second memory bank of the memory.
[0124] In Example 17, the subject matter of Examples 10-17 (excluding the present example) may optionally include dynamically increasing the power of at least one of a first storage body and a second storage body input to a memory based on a workload processed by a processor and a storage changing from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0125] Example 18 is an apparatus comprising means for performing a method as claimed in any preceding claim.
[0126] Example 19 is a system comprising:
[0127] a processor configured to process a workload having a noise tolerance level,
[0128] Memory, and
[0129] A controller is configured to select an input power level of the memory according to noise tolerance levels of workloads, wherein the workloads include a first workload having a first noise tolerance level and a second workload having a second noise tolerance level.
[0130] In Example 20, the subject matter of Examples 19-26 (excluding the present example) may optionally include wherein a second noise tolerance level for the second workload is lower than a first noise tolerance level for the first workload, and wherein a power level input to the memory associated with the second workload is higher than a power level input to the memory associated with the first workload being processed.
[0131] In Example 21, the subject matter of Examples 19-26 (excluding the present example) may optionally include, wherein the controller is configured to dynamically increase the power input to the memory based on a change in the workload of the processor and the memory from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0132] In Example 22, the subject matter of Examples 19-26 (excluding the present example) can optionally include, wherein the first workload is a first layer of the deep learning neural network and the second workload is a second, subsequent layer of the deep learning neural network.
[0133] In Example 23, the subject matter of Examples 19-26 (excluding the present example) can optionally include, wherein the first layer of the deep learning neural network is configured to process data having a first spatial correlation level, and the second layer of the deep learning neural network is configured to process data having a second spatial correlation level, the second spatial correlation level being weaker than the spatial correlation level of the data at the first spatial correlation level.
[0134] In Example 24, the subject matter of Examples 19-26 (excluding the present example) may optionally include, wherein the memory includes a first memory bank and a second memory bank, and the controller is further configured to store a data unit having a first subunit and a second subunit of data bits in the memory, including storing the first subunit of bits in the first memory bank of the memory, and storing the second subunit of bits in the second memory bank of the memory, wherein the bits of the first subunit are more efficient than the bits of the second subunit.
[0135] In Example 25, the subject matter of Examples 19-26 (excluding the present example) may optionally include, wherein the controller is configured to provide input power to a first memory body and a second memory body of the memory, wherein power is input to the first memory body of the memory at a higher power level than power input to the second memory body of the memory.
[0136] In Example 26, the subject matter of Examples 19-26 (excluding the present example) may optionally include, wherein the controller is further configured to dynamically increase the power of at least one of the first memory body and the second memory body input to the memory based on a change in the workload of the processor and the memory from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0137] Example 27 is an apparatus for use with a memory, comprising:
[0138] A memory controller having an input power mode selection logic module for selecting an input power level of a memory based on a noise tolerance level of a workload, wherein the workload includes a first workload having a first noise tolerance level and a second workload having a second noise tolerance level, and the memory controller also has a multi-stage power supply module for providing a selected power level to the memory.
[0139] In Example 28, the subject matter of Examples 27-34 (excluding the present example) may optionally include, wherein a second noise tolerance level for the second workload is lower than a first noise tolerance level for the first workload, and wherein a power level input to the memory associated with the second workload is higher than a power level input to the memory associated with the first workload.
[0140] In Example 29, the subject matter of Examples 27-34 (excluding the present example) may optionally include, wherein the input power mode selection logic module is further configured to dynamically increase the power input to the memory based on a change in the workload from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0141] In Example 30, the subject matter of Examples 27-34 (excluding the present example) can optionally include, wherein the first workload is a first layer of the deep learning neural network and the second workload is a second, subsequent layer of the deep learning neural network.
[0142] In Example 31, the subject matter of Examples 27-34 (excluding the present example) may optionally include, wherein the deep learning neural network has a first layer module for processing data having a first spatial correlation level, and a second layer module for processing data having a second spatial correlation level, the second spatial correlation level being weaker than the spatial correlation level of the data having the first spatial correlation level.
[0143] In Example 32, the subject matter of Examples 27-34 (excluding the present example) may optionally include, wherein the memory includes a first memory bank and a second memory bank, and wherein the memory controller has a data storage module for storing data units having first subunits and second subunits of data bits in the memory, including storing the first subunit of bits in the first memory bank of the memory, and storing the second subunit of bits in the second memory bank of the memory, wherein the bits of the first subunit are more efficient than the bits of the second subunit.
[0144] In Example 33, the subject matter of Examples 27-34 (excluding the present example) may optionally include, wherein the multi-stage power supply module of the memory controller is further configured to provide input power to a first memory bank and a second memory bank of the memory at a selected level, wherein the input power mode selection logic module is further configured to select the input power level of the first memory bank of the memory to be at a power level higher than the selected input power level of the second memory bank of the memory.
[0145] In Example 34, the subject matter of Examples 27-34 (excluding the present example) may optionally include, wherein the memory controller has a module for dynamically increasing the power of at least one of a first memory body and a second memory body input to the memory based on a change in the workload from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0146] Example 35 is a computer program product for a computing system, wherein the computer program product includes a computer-readable storage medium having program instructions embodied therewith, the program instructions executable by a processor of the computing system to cause operations including:
[0147] inputting power to the memory at a first power level,
[0148] processing a first workload having a first noise tolerance level,
[0149] processing a second workload having a second noise tolerance level different from the first noise tolerance level, and
[0150] Power is input to the memory at a second power level different from the first power level based on a noise tolerance level of a workload being processed.
[0151] In Example 36, the subject matter of Examples 35-42 (excluding the present example) may optionally include, wherein a second noise tolerance level for the second workload is lower than a first noise tolerance level for the first workload, and wherein a second power level of power input to the memory when the second workload is processed is higher than a first power level of power input to the memory when the first workload is processed.
[0152] In Example 37, the subject matter of Examples 35-42 (excluding the present example) may optionally include, wherein the operation further includes: dynamically increasing the power input to the memory from a first power level to a second level based on changing the workload processed by the processor and the memory from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0153] In Example 38, the subject matter of Examples 35-42 (excluding the present example) can optionally include, wherein the first workload is a first layer of the deep learning neural network and the second workload is a second, subsequent layer of the deep learning neural network.
[0154] In Example 39, the subject matter of Examples 35-42 (excluding the present example) can optionally include, wherein a first layer of the deep learning neural network processes data having a first spatial correlation level, and a second layer of the deep learning neural network processes data having a second spatial correlation level, the second spatial correlation level being weaker than the spatial correlation level of the data at the first spatial correlation level.
[0155] In Example 40, the subject matter of Examples 35-42 (excluding the present example) may optionally include, wherein the operation further includes: storing the data unit in a memory, wherein the storing includes storing a first sub-unit of bits in a first storage body of the memory, and storing a second sub-unit of bits in a second storage body of the memory, wherein the bits of the first sub-unit are more efficient than the bits of the second sub-unit.
[0156] In Example 41, the subject matter of Examples 35-42 (excluding the present example) may optionally include, wherein the operation further includes: inputting power to a first memory body and a second memory body of the memory, wherein the power is input to the first memory body of the memory at a higher power level than the power input to the second memory body of the memory.
[0157] In Example 42, the subject matter of Examples 35-42 (excluding the present example) may optionally include, wherein the operation further includes: dynamically increasing the power of at least one of the first storage body and the second storage body input to the memory based on a workload processed by the processor and the storage changing from a first workload having a first noise tolerance level to a second workload having a second noise tolerance level lower than the first noise tolerance level.
[0158] The described operations may be implemented as methods, devices or computer program products that use standard programming and / or engineering techniques to produce software, firmware, hardware or any combination thereof. The described operations may be implemented as computer program codes maintained in a "computer-readable storage medium", where a processor may read the code from a computer storage readable medium and execute the code. The computer-readable storage medium includes at least one of an electronic circuit, a storage material, an inorganic material, an organic material, a biological material, a housing, a shell, a coating and hardware. The computer-readable storage medium may include, but is not limited to, magnetic storage media (e.g., hard disk drives, floppy disks, magnetic tapes, etc.), optical storage devices (CD-ROMs, DVDs, optical disks, etc.), volatile and non-volatile memory devices (e.g., EEPROMs, ROMs, PROMs, RAMs, DRAMs, SRAMs, flash memories, firmware, programmable logic, etc.), solid-state devices (SSDs), etc. The code that implements the described operations may also be implemented in hardware logic implemented in hardware devices (e.g., integrated circuit chips, programmable gate arrays (PGAs), application-specific integrated circuits (ASICs), etc.). In addition, the code implementing the described operations can be implemented as a "transmission signal", where the transmission signal can be propagated through space or through a transmission medium such as optical fiber, copper wire, etc. The transmission signal in which the code or logic is encoded can also include wireless signals, satellite transmissions, radio waves, infrared signals, Bluetooth, etc. The program code embedded in the computer-readable storage medium can be sent from a sending station or computer to a receiving station or computer as a transmission signal. The computer-readable storage medium includes more than just transmission signals. Those skilled in the art will appreciate that many modifications can be made to this configuration without departing from the scope of this specification, and the product can include suitable information bearing media known in the art. Of course, those skilled in the art will appreciate that many modifications can be made to this configuration without departing from the scope of this specification, and the product can include any tangible information bearing medium known in the art.
[0159] In some applications, an apparatus according to the present description may be embodied in a computer system including a video controller, a device driver, and a network controller, the video controller causing information to be displayed on a monitor or other display coupled to the computer system, for example, the computer system including a desktop, workstation, server, mainframe, laptop, handheld computer, etc. Alternatively, an apparatus embodiment may be embodied in a computing device (e.g., a switch, router, etc.) that does not include a video controller, for example, or a computing device that does not include a network controller, for example.
[0160] The logic shown in the figure can show certain events that occur in a certain order. In alternative embodiments, certain operations can be performed, modified or removed in different orders. In addition, operations can be added to the above-mentioned logic and still meet the described embodiments. In addition, the operations described herein can occur sequentially or certain operations can be processed in parallel. In addition, operations can be performed by a single processing unit or by a distributed processing unit.
[0161] The above description of various embodiments has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit to the precise forms disclosed. Many modifications and variations are possible in light of the above teachings.
Claims
1. A device for use with a memory, comprising: a memory controller having an input power mode selection logic configured to select an input power level of the memory according to a noise tolerance level of a workload, wherein the workload includes a first workload having a first noise tolerance level and a second workload having a second noise tolerance level, and the memory controller further having a multi-stage power supply configured to provide the selected power level to the memory, wherein a second noise tolerance level of the second workload is lower than a first noise tolerance level of the first workload, and wherein a power level input to the memory associated with the second workload is higher than a power level input to the memory associated with the first workload, and Wherein, the first workload is a first layer of a deep learning neural network, and the second workload is a second, subsequent layer of the deep learning neural network.
2. The device according to claim 1, wherein: The input power mode selection logic is configured to dynamically increase power input to the memory according to a change in workload from the first workload having the first noise tolerance level to the second workload having a second noise tolerance level lower than the first noise tolerance level.
3. The device according to claim 1, wherein: The first layer of the deep learning neural network is configured to process data having a first level of spatial correlation, and the second layer of the deep learning neural network is configured to process data having a second level of spatial correlation, which is weaker than the spatial correlation level of the data at the first level of spatial correlation.
4. The device according to claim 1, wherein: The memory includes a first memory bank and a second memory bank, and wherein the memory controller is configured to store data units having first sub-units and second sub-units of data bits in the memory, including storing the first sub-unit of bits in the first memory bank of the memory, and storing the second sub-unit of bits in the second memory bank of the memory, wherein the bits of the first sub-unit are more efficient than the bits of the second sub-unit.
5. The device according to claim 4, wherein: The multi-stage power supply of the memory controller is configured to provide input power to a first memory bank and a second memory bank of the memory at a selected level, wherein the input power mode selection logic is configured to select the input power level of the first memory bank of the memory to be at a power level higher than the selected input power level of the second memory bank of the memory.
6. The device according to claim 5, wherein: The memory controller is also configured to dynamically increase power input to at least one of the first memory bank and the second memory bank of the memory based on a change in workload from the first workload having the first noise tolerance level to the second workload having the second noise tolerance level lower than the first noise tolerance level.
7. A method for a memory, comprising: inputting power to the memory at a first power level; Using the processor and the memory, processing a first workload having a first noise tolerance level; processing, using the processor and the memory, a second workload having a second noise tolerance level different from the first noise tolerance level; as well as inputting power to the memory at a second power level different from the first power level according to a noise tolerance level of a workload being processed, wherein the second noise tolerance level of the second workload is lower than the first noise tolerance level of the first workload, and wherein the second power level of power input to the memory when the second workload is processed is higher than the first power level of power input to the memory when the first workload is processed, and Wherein, the first workload is a first layer of a deep learning neural network, and the second workload is a second, subsequent layer of the deep learning neural network.
8. The method according to claim 7 further includes dynamically increasing the power input to the memory from the first power level to the second power level based on changing the workload processed by the processor and the memory from the first workload having the first noise tolerance level to the second workload having the second noise tolerance level lower than the first noise tolerance level.
9. The method according to claim 7, wherein: The first layer of the deep learning neural network processes data having a first spatial correlation level, and the second layer of the deep learning neural network processes data having a second spatial correlation level, which is weaker than the spatial correlation level of the data at the first spatial correlation level.
10. The method of claim 7, further comprising storing a data unit in the memory, wherein: Storing includes storing a first subunit of bits in a first bank of the memory and storing a second subunit of bits in a second bank of the memory, wherein the first subunit of bits is more significant than the second subunit of bits.
11. The method of claim 10, further comprising inputting power to the first memory bank and the second memory bank of the memory, wherein: Power is input to a first bank of the memory at a higher power level than power input to a second bank of the memory.
12. The method according to claim 11 further includes dynamically increasing the power input to at least one of the first memory body and the second memory body of the memory according to the workload processed by the processor and the memory changing from the first workload having the first noise tolerance level to the second workload having the second noise tolerance level lower than the first noise tolerance level.
13. A computing system comprising: a processor configured to process a workload having a noise tolerance level; Memory; as well as a controller configured to select an input power level of the memory according to a noise tolerance level of a workload, wherein the workload includes a first workload having a first noise tolerance level and a second workload having a second noise tolerance level, wherein a second noise tolerance level of the second workload is lower than a first noise tolerance level of the first workload, and wherein a power level input to the memory associated with the second workload is higher than a power level input to the memory associated with the first workload being processed, and Wherein, the first workload is a first layer of a deep learning neural network, and the second workload is a second, subsequent layer of the deep learning neural network.
14. The computing system of claim 13, wherein: The controller is configured to dynamically increase power input to the memory according to a change in a workload of the processor and the memory from the first workload having the first noise tolerance level to the second workload having a second noise tolerance level lower than the first noise tolerance level.
15. The computing system of claim 13, wherein: The first layer of the deep learning neural network is configured to process data having a first level of spatial correlation, and the second layer of the deep learning neural network is configured to process data having a second level of spatial correlation, which is weaker than the spatial correlation level of the data at the first level of spatial correlation.
16. The computing system of claim 13, wherein: The memory includes a first memory bank and a second memory bank, and the controller is further configured to store a data unit having a first subunit and a second subunit of data bits in the memory, including storing the first subunit of bits in the first memory bank of the memory, and storing the second subunit of bits in the second memory bank of the memory, wherein the bits of the first subunit are more efficient than the bits of the second subunit.
17. The computing system of claim 16, wherein: The controller is configured to provide input power to a first memory bank and a second memory bank of the memory, wherein power is input to the first memory bank of the memory at a higher power level than power input to the second memory bank of the memory.
18. The computing system of claim 17, wherein: The controller is also configured to dynamically increase power input to at least one of a first memory body and a second memory body of the memory based on a change in a workload of the processor and the memory from the first workload having the first noise tolerance level to the second workload having a second noise tolerance level lower than the first noise tolerance level.
19. A computer-readable medium having computer-executable instructions stored thereon, for causing a computer system to execute the method according to any one of claims 7 to 12.
20. A computer program product having computer executable instructions for causing a computer system to execute the method according to any one of claims 7 to 12.
