Neural network acceleration device and control method thereof

By introducing signal detectors and processing elements into neural network acceleration devices, unnecessary arithmetic operations and memory read operations are skipped, and problems of high power consumption and slow data processing speed in the prior art are solved, achieving more efficient AI task execution.

CN111723922BActive Publication Date: 2024-11-26SK HYNIX INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201910982799.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-03-20
Filing Date
2019-10-16
Publication Date
2024-11-26
Estimated Expiration
2039-10-16

AI Technical Summary

Technical Problem

When performing AI tasks, existing neural network acceleration devices have high power consumption and low data processing speed due to frequent arithmetic operations and memory read operations.

Method used

An acceleration device is designed, including a signal detector and processing element (PE) to generate a final zero-value flag signal, a first-value flag signal and a second-value flag signal by converting the input signal into a flag signal, based on these The flag signal skips unnecessary arithmetic operations and memory read operations, thereby reducing dynamic current and power consumption.

Benefits of technology

It effectively reduces the power consumption of the computing unit, especially in binary neural networks by skipping operations more frequently, significantly reduces power consumption and improves data processing speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111723922B_ABST
    Figure CN111723922B_ABST
Patent Text Reader

Abstract

The present application relates to an acceleration device. The acceleration device includes a signal detector, which converts a first input signal and a second input signal into a first converted input signal and a second converted input signal, respectively, and generates a final zero-value flag signal, a first one-value flag signal, and a second one-value flag signal. The acceleration device further includes a processing element (PE), which processes the first converted input signal and the second converted input signal based on the final zero-value flag signal, the first one-value flag signal, and the second one-value flag signal, and skips the first arithmetic operation and the second arithmetic operation when the final zero-value flag signal has a first value. The first value of the final zero-value flag signal indicates that the first input signal or the second input signal or both have a value of 0.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of Korean Patent Application No. 10-2019-0031611 filed on March 20, 2019, in the Korean Intellectual Property Office, which is hereby incorporated by reference in its entirety. Technical Field

[0003] The present disclosure relates to an acceleration device and a processing system including the acceleration device, as well as a method for controlling the acceleration device. In particular, some embodiments relate to a neural network acceleration device. Background Art

[0004] Various artificial neural networks (ANNs) are effective tools for performing artificial intelligence (AI) tasks. For example, deep neural networks (DNNs) are used for AI tasks including computer vision, object and speech recognition, medical diagnosis, game playing, and robotics. DNNs include convolutional neural networks (CNNs) with multiple convolutional layers, each of which generates a successive higher level of abstraction of the input data. DNNs also include binary neural networks (BNNs) with binary weights.

[0005] A neural network accelerator device (NNAD) is a microprocessor or computer system designed to perform specific AI tasks. For example, a NNAD may include multiple processing elements (PEs) that perform multiply-accumulate (MAC) operations in parallel. Each MAC operation requires multiple read operations and multiple arithmetic operations to the memory (e.g., off-chip DRAM), which may increase power consumption and reduce data processing speed. Summary of the invention

[0006] Embodiments of the present application relate to an acceleration device, a processing system including the acceleration device, and a method for controlling the acceleration device.

[0007] In an embodiment, the acceleration device includes a signal detector that converts a first input signal and a second input signal into a first converted input signal and a second converted input signal, respectively, and generates a final zero-value flag signal, a first one-value flag signal, and a second one-value flag signal. The acceleration device further includes a processing element (PE) that processes the first converted input signal and the second converted input signal based on the final zero-value flag signal, the first one-value flag signal, and the second one-value flag signal, and skips the first arithmetic operation and the second arithmetic operation when the final zero-value flag signal has a first value. The first value of the final zero-value flag signal indicates that the first input signal or the second input signal or both have a value of 0.

[0008] In an embodiment, a processing system includes a memory and an acceleration device, wherein the memory is configured to store data and output an input signal in response to a request signal from a host, and the acceleration device receives the input signal and performs one or more given arithmetic operations on the input signal. The input signal includes a first input signal and a second input signal. The acceleration device includes a signal detector, which converts the first input signal and the second input signal into a first converted input signal and a second converted input signal, respectively, and generates a final zero value flag signal, a first one-value flag signal, and a second one-value flag signal. The acceleration device further includes a processing element (PE), which processes the first converted input signal and the second converted input signal based on the final zero value flag signal, the first one-value flag signal, and the second one-value flag signal, and when the final zero value flag signal has a first value, skips the first arithmetic operation and the second arithmetic operation. The first value of the final zero value flag signal indicates that the first input signal or the second input signal or both have a value of 0.

[0009] In an embodiment, a method of controlling an acceleration device includes: converting a first input signal and a second input signal into a first converted input signal and a second converted input signal, respectively; generating a final zero-value flag signal, a first one-value flag signal, and a second one-value flag signal; and when the final zero-value flag signal has a first value, skipping a first arithmetic operation and a second arithmetic operation. The first value of the final zero-value flag signal indicates that the first input signal or the second input signal or both have a value of 0. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A processing system and a host according to an embodiment of the present disclosure are shown.

[0011] Figure 2 A neural network acceleration device (NNAD) according to an embodiment of the present disclosure is shown.

[0012] Figure 3 A signal detector and a PE cluster according to an embodiment of the present disclosure are shown.

[0013] Figure 4 A signal detector according to an embodiment of the present disclosure is shown.

[0014] Figure 5 The values ​​of the first input signal, the second input signal, the first converted input signal, the second converted input signal, the first zero-value flag signal, the second zero-value flag signal, the final zero-value signal, the first one-value flag signal, and the second one-value flag signal are shown according to an embodiment.

[0015] Figure 6 A PE according to an embodiment of the present disclosure is shown.

[0016] Figure 7 A PE according to an embodiment of the present disclosure is shown.

[0017] Figure 8 are the values ​​of the first input signal, the second input signal, the first converted input signal, the second converted input signal, the first zero-value flag signal, the second zero-value flag signal, the final zero-value signal, the first one-value flag signal, the second one-value flag signal, the first control signal, the second control signal, and the third control signal according to an embodiment.

[0018] Fig. 9 is a flowchart illustrating a process performed by a neural network acceleration device according to an embodiment.

[0019] Fig.10 is a flow chart illustrating a process of converting a first input signal and a second input signal into a first converted input signal and a second converted input signal, respectively, according to an embodiment.

[0020] Fig.11 is a flow chart illustrating a process for generating one or more flag signals according to an embodiment.

[0021] Fig.12 is a flowchart illustrating a process of controlling one or more read operations of a first internal memory and one or more operations of an operation unit according to an embodiment. DETAILED DESCRIPTION

[0022] Embodiments of the present application relate to an acceleration device and a processing system including the acceleration device, as well as a method for controlling the acceleration device. The acceleration device includes a processing element (PE), which skips a first arithmetic operation and a second arithmetic operation when a final zero-value flag signal has a first value. When the first one-value flag signal or the second one-value flag signal or both have a first value, and the final zero-value flag signal has a second value, the PE further skips the first arithmetic operation and performs the second arithmetic operation. By skipping the first arithmetic operation and the second arithmetic operation, or skipping the first arithmetic operation, one or more dynamic currents are substantially prevented from appearing in a computing unit of the PE, thereby reducing the power consumption of the computing unit.

[0023] In an embodiment, the acceleration device further includes a signal detector configured to convert the first input signal and the second input signal into a first converted input signal and a second converted input signal, respectively, and generate a final zero-value flag signal, a first one-value flag signal, and a second one-value flag signal. When the acceleration device is used for a binarized neural network (BNN), the acceleration device can reduce power consumption by skipping one or more operations more frequently than for other types of neural networks.

[0024] In an embodiment, the PE includes a controller configured to skip a first read operation of a value of a first converted input signal from a first internal memory and a second read operation of a value of a second converted input signal from a second internal memory when the final zero-value flag signal has a first value. When the first one-value flag signal or the second one-value flag signal or both have a first value and the final zero-value flag signal has a second value, the controller further skips the first read operation or the second read operation. By skipping the first read operation or the second read operation or both, the number of times the first internal memory or the second internal memory or both are accessed using one or more signals (e.g., a chip enable signal, an address signal) can be reduced, thereby reducing power consumption.

[0025] A detailed description of the embodiments is provided below in conjunction with the accompanying drawings. The scope of the present disclosure is limited only by the claims, and many alternatives, modifications, and equivalents are contemplated. Although the steps of the various processes are presented in a given order, the embodiments are not necessarily limited to being performed in the order listed. In some embodiments, certain operations may be performed simultaneously, in an order different from the order described, or not performed at all.

[0026] In the following description, many specific details are set forth. These details are provided to facilitate a thorough understanding of the scope of the present disclosure through specific examples, and embodiments may be practiced according to the claims without some of these specific details. Therefore, the specific embodiments of the present disclosure are illustrative, not exclusive or restrictive. For the purpose of clarity, the technical content known in the technical field related to the present disclosure is not described in detail to avoid unnecessarily obscuring the present disclosure.

[0027] Figure 1 A processing system 100 and a host 102 according to an embodiment are shown. The processing system 100 includes a high-speed interface 120, a first memory 130, a memory interface 140, a bus interface 150, a second memory 160, and a neural network acceleration device (NNAD) 110.

[0028] The high-speed interface 120 communicates with the host 102. In an embodiment, the high-speed interface 120 is implemented using one of a Peripheral Component Interconnect Express (PCIe) protocol, an Advanced Technology Attachment (ATA) protocol, a Serial ATA (SATA) protocol, a Parallel ATA (PATA) protocol, a Small Computer Small Interface (SCSI) protocol, a Serial Attached SCSI (SAS) protocol, a Universal Serial Bus (USB) protocol, a MultiMediaCard (MMC) protocol, an Enhanced MiniDisk Interface (ESDI) protocol, and an Integrated Drive Electronics (IDE) protocol.

[0029] The bus interface 150 interfaces with the high-speed interface 120, the second memory 160, the NNAD 110, and the memory interface 140. The memory interface 140 interfaces with the first memory 130.

[0030] The first memory 130 stores data transmitted from the host 102 or the NNAD 110. In an embodiment, the first memory 130 is a dynamic random access memory (DRAM). However, the embodiments of the present disclosure are not limited thereto.

[0031] The second memory 160 stores data to be transmitted to the NNAD 110. In an embodiment, the second memory 160 is a static random access memory (SRAM). However, the embodiments of the present disclosure are not limited thereto, and the second memory 160 may be a type other than SRAM or DRAM, for example, a next generation memory such as MRAM, F-RAM, RRAM, etc.

[0032] The second memory 160 provides an input signal IS to the NNAD 110 in response to a request signal from the host 102. For example, the request signal may be transmitted from the host 102 to the second memory 160 through the high-speed interface 120 and the bus interface 150. In addition, the request signal may be transmitted from the host 102 to the first memory 130 through the high-speed interface 120, the bus interface 150, and the memory interface 140, and then transmitted from the first memory 130 to the second memory 160 through the memory interface 140 and the bus interface 150.

[0033] NNAD 110 receives an input signal IS, performs one or more given operations on the input signal IS, and provides an output value to host 102 through bus interface 150 and high-speed interface 120, or provides an output value to first memory 130 through memory interface 140. In an embodiment, NNAD 210 includes a processing element (PE) cluster (or PE array).

[0034] Figure 2 A method suitable for use as a Figure 1 NNAD 210 of NNAD 110 in . NNAD 210 includes a signal detector 230, a PE cluster 250 and an output feature generator 270.

[0035] The signal detector 230 converts the one or more input signals IS into one or more converted input signals CIS, respectively, and generates one or more flag signals FS, each flag signal FS indicating that the one or more converted input signals CIS have a specific value (e.g., 0 or 1). The signal detector 230 provides the converted input signals CIS and the flag signals FS to the PE cluster 250.

[0036] The PE cluster 250 processes the converted input signal CIS based on one or more flag signals FS. In an embodiment, based on the one or more flag signals FS, the PE cluster 250 may perform a first arithmetic operation (e.g., a multiplication operation) on the converted input signal CIS, and perform a second arithmetic operation (e.g., an addition operation) on the result of the first arithmetic operation and the result of the previous addition operation, or skip the first arithmetic operation, or skip the first arithmetic operation and the second arithmetic operation. For example, the PE cluster 250 may include a plurality of processing elements (PEs) that perform MAC operations in parallel, and each of the PEs may include an operation unit and one or more internal memories (e.g., a scratch pad memory, a register file).

[0037] Output feature generator 270 receives one or more output signals OS from PE cluster 250, applies an enabling function to the value of output signal OS to convert the value into a non-linear value, performs aggregation (e.g., maximum aggregation) on the non-linear value, and transmits the aggregated value to a memory (e.g., Figure 1 The first memory 130 or the second memory 160 in the host computer (eg, Figure 1 host 102), or transmitted to both.

[0038] Figure 3 A signal detector 330 is shown having a signal converter 312 and a flag generation circuit 320 according to an embodiment of the present disclosure.

[0039] The signal converter 312 receives one or more input signals IS and generates one or more converted input signals CIS in response to the received input signals IS. In an embodiment, the input signal IS includes a first input signal and a second input signal, and the signal converter 312 converts the first input signal and the second input signal into a first converted input signal and a second converted input signal, respectively. The signal converter 312 provides the converted input signal CIS to the flag generation circuit 320 and the PE cluster 350.

[0040] The flag generation circuit 320 generates one or more flag signals FS in response to the converted input signal CIS, and provides the flag signal FS to the PE cluster 350. For example, the flag signal FS includes a zero-value flag signal indicating that the first converted input signal or the second converted input signal or both have a value of 0. The flag signal FS further includes a first one-value flag signal indicating whether the first converted input signal has a value of 1, and a second one-value flag signal indicating whether the second converted input signal has a value of 1.

[0041] Figure 4The signal detector 430 having a signal converter 412 and a flag generation circuit 420 is shown. According to an embodiment of the present disclosure, the flag generation circuit 420 includes a first flag generator 414, a second flag generator 416 and a lookup table (LUT) 418. Figure 5 The operation of the signal detector 430 is described in more detail. Figure 5 The values ​​of the first input signal Feature, the second input signal Weight, the first converted input signal cf, the second converted input signal cw, the first zero-value flag signal zff, the second zero-value flag signal zfw, the final zero-value signal zf, the first one-value flag signal off and the second one-value flag signal ofw according to the embodiment are shown.

[0042] The signal converter 412 receives the first input signal Feature and the second input signal Weight, and generates the first converted input signal cf and the second converted input signal cw in response to the first input signal Feature and the second input signal Weight. Under a first condition when the first input signal Feature or the second input signal Weight or both have a zero value, the signal converter 412 generates the first converted input signal cf and the second converted input signal cw having the same value as the first input signal Feature and the second input signal Weight, respectively. For example, referring to Figure 5 , when the first input signal Feature and the second input signal Weight have values ​​(0, -1), (-1, 0), (0, 0), (1, 0), (any other integer, 0), (-(any other integer), 0), (0, 1), (0, any other integer), and (0, -(any other integer)), the first converted input signal cf and the second converted input signal cw have the same values ​​(0, -1), (-1, 0), (0, 0), (1, 0), (any other integer, 0), (-(any other integer), 0), (0, 1), (0, any other integer), and (0, -(any other integer)). The value of any other integer indicates an integer value other than 1, 0, and -1.

[0043] Under the second condition, the second input signal Weight has a value of -1 or 1. In an embodiment, the first condition has a higher priority than the second condition, so the first input signal Feature does not have a value of 0 under the second condition. When the second input signal Weight has a value of -1, the signal converter 412 generates a first converted input signal cf that is a reverse sign version of the first input signal Feature, and a second converted input signal cw having a value of 1. For example, referring to Figure 5, when the first input signal Feature and the second input signal Weight have values ​​(-1, -1), (1, -1), (any other integer, -1), (-(any other integer), -1), the first converted input signal cf and the second converted input signal cw have values ​​(1, 1), (-1, 1), (-(any other integer), 1), ((any other integer), 1), respectively. When the second input signal Weight has a value of 1, the signal converter 412 generates a first converted input signal cf having the same value as the first input signal Feature and a second converted input signal cw having a value of 1. For example, referring to Figure 5 , when the first input signal Feature and the second input signal Weight have values ​​(-1,1), (1,1), (any other integer, 1) and (-(any other integer), 1), the first converted input signal cf and the second converted input signal cw have values ​​(-1,1), (1,1), (any other integer, 1) and (-(any other integer), 1) respectively.

[0044] Under the second condition, the second converted input signal cw has a value of 1. Therefore, unless the value is stored in a PE (e.g., a chip enable signal, an address signal) that needs to generate one or more signals (e.g., a chip enable signal, an address signal) to access the memory, Figure 7 PE760-1 in the specific memory (for example, Figure 7 When the PE has a high-speed scratchpad memory 760-2 in the memory and can know the information about the value of the second converted input signal cw (ie, 1), otherwise, the PE can skip the read operation on the memory to reduce power consumption.

[0045] Under the second condition, the second converted input signal cw has a value of 1, regardless of whether the second input signal Weight has a value of 1 or -1. In an embodiment, a binary neural network (BNN) may use -1 / 1 or -1 / 0 / 1 as the value of the second input signal Weight. When a neural network acceleration device (e.g., a PE cluster) including a PE cluster according to an embodiment of the present disclosure is used, Figure 2 When the NNAD 210 in FIG. 1 is used in a BNN, such a NNAD may skip one or more operations (e.g., read operations, multiplication operations) in response to the values ​​of the second input signal Weight of -1 and 1. Therefore, power consumption may be reduced when the NNAD is used in a BNN compared to other types of neural networks.

[0046] Under the third condition, the first input signal Feature has a value of -1 or 1. In an embodiment, the first condition has a higher priority than the second condition, and the second condition has a higher priority than the third condition, so under the third condition, the second input signal Weight has a value other than 0, 1, and -1. When the first input signal Feature has a value of -1, the signal converter 412 generates a first converted input signal cf having a value of 1, and a second converted input signal cw which is a reverse sign version of the second input signal Weight. For example, referring to Figure 5 , when the first input signal Feature and the second input signal Weight have values ​​(-1, any other integer) and (-1, -(any other integer)), the first converted input signal cf and the second converted input signal cw have values ​​(1, -(any other integer)) and (1, (any other integer)), respectively. When the first input signal Feature has a value of 1, the signal converter 412 generates a first converted input signal cf having a value of 1 and a second converted input signal cw having the same value as the second input signal Weight. For example, referring to Figure 5 , when the first input signal Feature and the second input signal Weight have values ​​(1, any other integer) and (1, -(any other integer)), the first converted input signal cf and the second converted input signal cw have values ​​(1, any other integer) and (1, -(any other integer)), respectively.

[0047] Under the fourth condition, each of the first input signal Feature and the second input signal Weight has any other integer or -(any other integer), and the signal converter 412 generates the first converted input signal cf and the second converted input signal cw having the same value as the first input signal Feature and the second input signal Weight. Figure 5 , when the first input signal Feature and the second input signal Weight have values ​​of (any other integer, any other integer), (-(any other integer), any other integer), (any other integer, -(any other integer)) and (-(any other integer), -(any other integer)), the first converted input signal cf and the second converted input signal cw have values ​​of (any other integer, any other integer), (-(any other integer), any other integer), (any other integer, -(any other integer)) and (-(any other integer), -(any other integer)) respectively.

[0048] exist Figure 5In the illustrated embodiment, the priority is given in the order of the first, second, third and fourth conditions. For example, when a pair of values ​​of the first input signal Feature and the second input signal Weight satisfies two or more of the first, second, third and fourth conditions, the signal converter 412 generates the first converted input signal cf and the second converted input signal cw according to the condition with the highest priority. However, the embodiments of the present disclosure are not limited thereto.

[0049] The signal converter 412 provides the generated first converted input signal cf to the first flag generator 414 and the PE cluster (eg, Figure 2 The signal converter 412 provides the generated second converted input signal cw to the second flag generator 416 and the PE cluster.

[0050] The first flag generator 414 receives the first converted input signal cf, and generates a first zero value flag signal zff and a first one value flag signal off in response to the first converted input signal cf. In an embodiment, when the first converted input signal cf has a value of 0, the first flag generator 414 generates a first zero value flag signal zff having a first value (e.g., a logical high value) and a first one value flag signal off having a second value (e.g., a logical low value). In addition, when the first converted input signal cf has a value other than 0, the first flag generator 414 generates a first zero value flag signal zff having a second value and a first one value flag signal off having a first value or a second value. For example, when the first converted input signal cf has a value of 1, the first flag generator 414 generates a first zero value flag signal zff having a second value and a first one value flag signal off having a first value. When the first converted input signal cf has a value other than 0 and 1, the first flag generator 414 may generate a first zero value flag signal zff having a second value and a first one value flag signal off having a second value. In Figure 5 In the illustrated embodiment, the first value is 1, and the second value is 0. However, the embodiments of the present disclosure are not limited thereto.

[0051] The second flag generator 416 receives the second converted input signal cw and generates a second zero-value flag signal zfw and a second one-value flag signal ofw in response to the second converted input signal cw. In an embodiment, the second flag generator 416 operates similarly to the first flag generator 414 (asserting the second zero-value flag signal zfw to indicate that the second converted input signal cw has a zero value, and asserting the second one-value flag signal ofw to indicate that the second converted input signal cw has a one value), so for the sake of brevity, a detailed description thereof is omitted.

[0052] LUT 418 generates a third zero-value flag signal (or final zero-value flag signal) zf in response to the first zero-value flag signal zff and the second zero-value flag signal zfw, and provides the third zero-value flag signal zf to the PE cluster. In an embodiment, the third zero-value flag signal zf is the result of a logical OR operation of the first zero-value flag signal zff and the second zero-value flag signal zfw. For example, referring to Figure 5 , when the first zero-value flag signal zff or the second zero-value flag signal zfw or both have a value of 1, the LUT 418 generates a third zero-value flag signal zf having a first value. Otherwise, the LUT 418 generates a third zero-value flag signal zf having a second value. Figure 5 In the illustrated embodiment, the first value is 1, and the second value is 0. However, the embodiments of the present disclosure are not limited thereto.

[0053] Figure 6 The embodiment according to the present disclosure is suitable for use as Figure 3 A processing element (PE) 650-1 is included in a PE cluster 350 of the PE cluster 350. The PE 650-1 includes a first internal memory array 660, a second internal memory array 670, a controller (eg, PE controller) 680, and an operation unit 690.

[0054] The first internal memory array 660 stores the value of the converted input signal CIS. In an embodiment, the first internal memory array 660 includes a plurality of memories, each of which is a first type of memory and requires the generation of one or more signals (e.g., a chip enable signal, an address signal) to access the data stored therein. For example, the first internal memory array 660 may include a plurality of high-speed scratch pad memories.

[0055] The second internal memory array 670 stores the value of the flag signal FS. In an embodiment, the second internal memory array 670 includes a plurality of memories, each of which is a second type memory different from the first type memory. For example, the second memory 670 may include a plurality of register files (or registers).

[0056] The controller 680 controls one or more read operations to the plurality of memories of the first internal memory array 660 based on one or more values ​​stored in the second internal memory array 670. The PE controller 680 further controls one or more arithmetic operations of the operation unit 690 based on the values ​​stored in the second internal memory array 670.

[0057] Figure 7 The embodiment according to the present disclosure is suitable for use as Figure 6The processing element (PE) 750-1 of the PE 650-1 in FIG. PE 750-1 includes a first internal memory array 760, a second internal memory array 770, a controller (eg, a PE controller) 780, and an operation unit (or operation circuit) 790. Figure 7 In the illustrated embodiment, PE 750-1 is configured to perform a convolution operation. However, the embodiments of the present disclosure are not limited thereto.

[0058] The first internal memory array 760 includes a first high-speed scratchpad memory 760-1 and a second high-speed scratchpad memory 760-2. The first high-speed scratchpad memory 760-1 stores the value of the first converted input signal cf, and the second high-speed scratchpad memory 760-2 stores the value of the second converted input signal cw. Each of the first high-speed scratchpad memory 760-1 and the second high-speed scratchpad memory 760-2 can be used as a temporary storage device to save data for fast retrieval in PE 750-1. Therefore, it is not necessary to store data from an external memory (e.g., Figure 1 Compared with the first memory 130 in the memory 130, the power consumption and delay of transmitting the data stored in each of the first high-speed scratch pad memory 760-1 and the second high-speed scratch pad memory 760-2 can be reduced. In an embodiment, each of the first high-speed scratch pad memory 760-1 and the second high-speed scratch pad memory 760-2 is a static random access memory (SRAM), but the embodiments of the present disclosure are not limited thereto. One or more signals (e.g., a chip enable signal, an address signal) can be applied to access the data stored in each of the first high-speed scratch pad memory 760-1 and the second high-speed scratch pad memory 760-2.

[0059] The second internal memory array 770 includes a first register array 770-1, a second register array 770-2, and a third register array 770-3. The first register array 770-1 stores a zero-value flag signal zf (e.g., Figure 4 The second register array 770-2 stores the value of the third zero-value signal zf in the first register array 770-2, and the first one-value flag signal off (for example, Figure 4 The value of the first one-valued flag signal off) in the third register array 770-3 is stored, and the second one-valued flag signal ofw (for example, Figure 4 The value of the second one-value flag signal ofw) in.

[0060] The PE controller 780 controls the read operation of the first scratch pad memory 760-1 and the second scratch pad memory 760-2 based on one or more values ​​stored in the first, second, and third register arrays 770-1 to 770-3. When the PE controller 780 determines that the value of the zero-value flag signal zf stored in the first register array 770-1 is the first value (e.g., 1), the PE controller 780 controls the PE 750-1 to skip the read operation of the first scratch pad memory 760-1 and the read operation of the second scratch pad memory 760-2. For example, referring to Figure 8 When the values ​​of the zero-value flag signal zf, the first one-value flag signal off, and the second one-value flag signal ofw are (1, 0, 0) as shown in the solid-line box, the PE controller 780 controls the PE 750-1 to skip the read operation on the first high-speed scratch pad memory 760-1 and the read operation on the second high-speed scratch pad memory 760-2.

[0061] When the PE controller 780 determines that the value of the zero-value flag signal zf stored in the first register array 770-1 is the second value (e.g., 0) and one or both of the values ​​of the first one-value flag signal off and the second one-value flag signal ofw are the first value (e.g., 1), the PE controller 780 controls the PE 750-1 to skip the read operation of the first high-speed scratch pad memory 760-1 or the read operation of the second high-speed scratch pad memory 760-2. Figure 8 When the values ​​of the zero-value flag signal zf, the first single-value flag signal off, and the second single-value flag signal ofw are (0, 0, 1) as shown in the dashed box, the PE controller 780 controls the PE 750-1 to skip the read operation on the second high-speed scratch pad memory 760-2. In addition, when the values ​​of the zero-value flag signal zf, the first single-value flag signal off, and the second single-value flag signal ofw are (0, 1, 0) as shown in the dashed box, the PE controller 780 controls the PE 750-1 to skip the read operation on the first high-speed scratch pad memory 760-1.

[0062] As described above, the PE controller 780 may control the PE 750-1 to skip one or both of the read operation to the first scratch pad memory 760-1 and the read operation to the second scratch pad memory 760-2 based on the values ​​stored in the first, second, and third register arrays 770-1 to 770-3. Therefore, the number of times the first scratch pad memory 760-1, the second scratch pad memory 760-2, or both are accessed using one or more signals (e.g., a chip enable signal, an address signal) may be reduced, thereby reducing power consumption.

[0063] The PE controller 780 further controls the multiplication and addition operations (e.g., accumulation operations) of the operation unit 790 based on one or more values ​​stored in the first, second, and third register arrays 770-1 to 770-3. The operation unit 790 includes a multiplication unit (or multiplication circuit) 736 and an addition unit (or addition circuit) 738. The multiplication unit 736 includes first and second flip-flops (FFs) 722 and 724, a multiplier 726, and a multiplexer (MUX) 728. The addition unit 738 includes an adder 732 and a third FF 734.

[0064] In an embodiment, the PE controller 780 may access the first, second, and third register arrays 770-1, 770-2, and 770-3 in the order of a given priority. For example, the PE controller 780 may first access the first register array 770-1 having the highest priority among the register arrays 770-1, 770-2, and 770-3. When the PE controller 780 determines that the value of the zero-value flag signal zf stored in the first register array 770-1 is the first value (e.g., 1), the PE controller 780 generates a first control signal c0 to disable the first and second flip-flops 722 and 724, generates a second control signal c1 to control the MUX 728 to keep outputting the previous multiplication result, and generates a third control signal c2 to disable the third flip-flop 734.

[0065] For example, refer to Figure 8 , when the values ​​of the zero-value flag signal zf, the first one-value flag signal off, and the second one-value flag signal ofw are (1, 0, 0) as shown in the solid-line box, the PE controller 780 generates a first control signal c0 having a value of 0, a second control signal c1 having a value of 1, and a third control signal c2 having a value of 0. Because the multiplier 726 keeps receiving the previous input values ​​from the first and second FFs 722 and 724, and the adder 732 keeps receiving the previous multiplication result from the MUX 728 and the previous value from the third FF 734, the operation unit 790 can skip the multiplication operation and the addition operation. By skipping the multiplication operation and the addition operation, one or more dynamic currents are substantially prevented from occurring in the multiplier 726 and the adder 732, thereby reducing the power consumption of the operation unit 790.

[0066] When the PE controller 780 determines that the value of the zero-value flag signal zf stored in the first register array 770-1 is the second value (e.g., 0), the PE controller 780 may access the second register array 770-2 or the third register array 770-3, or both. When the PE controller 780 determines that the value of the zero-value flag signal zf stored in the first register array 770-1 is the second value (e.g., 0) and one or both of the values ​​of the first one-value flag signal off and the second one-value flag signal ofw are the first value (e.g., 1), the PE controller 780 generates a first control signal c0 to disable the first and second flip-flops 722 and 724, generates a second control signal c1 to control the MUX 728 to bypass data from the first or second scratch pad memory 760-1 or 760-2 to the adder 732, and generates a third control signal c3 to enable the third FF 734. In particular, when the value of the zero-value flag signal zf is 0 and the value of the second one-value flag signal ofw is 1, the second control signal c1 controls the MUX 728 to output the Feature value from the first scratch pad memory 760-1 to the adder 732, and when the zero-value flag signal zf is 0, the first one-value flag signal off is 1, and the second one-value flag signal ofw is 0, the second control signal c1 controls the MUX 728 to output the Weight value from the second scratch pad memory 760-2 to the adder 732. In this way, the PE 750 avoids incurring the power consumption cost of using the multiplier 760 to perform a simple multiplication by 1.

[0067] For example, refer to Figure 8 , when the values ​​of the zero-value flag signal zf, the first one-value flag signal off, and the second one-value flag signal ofw are (0, 0, 1) as shown in the dashed box, the PE controller 780 generates the first control signal c0 having a value of 0, the second control signal c1 having a value of 0, and the third control signal c2 having a value of 1. In another example, when the values ​​of the zero-value flag signal zf, the first one-value flag signal off, and the second one-value flag signal ofw are (0, 1, 0) as shown in the dashed box, the PE controller 780 generates the first control signal c0 having a value of 0, the second control signal c1 having a value of 2, and the third control signal c2 having a value of 1. In both examples, because the multiplier 726 keeps receiving the previous input values ​​from the first and second FFs 722 and 724, and the adder 732 receives the current value from the second cache memory 760-2 and receives the current value from the third FF 734, the operation unit 790 can skip the multiplication operation and perform the addition operation. By skipping the multiplication operation, one or more dynamic currents in the multiplier 726 are substantially prevented from occurring, thereby reducing power consumption of the arithmetic unit 790 .

[0068] exist Figure 7In the illustrated embodiment, when the first control signal c0 is not determined to have a value of 0, the first and second FFs 722 and 724 are disabled, so the multiplication unit 736 skips the multiplication operation of the first converted input signal cf and the second converted input signal cw, and outputs the previous multiplication result. However, the embodiments of the present disclosure are not limited thereto. For example, when it is determined that the first control signal (or multiplication input hold signal) c0 has a value of 1, the first and second FFs 722 and 724 are disabled.

[0069] exist Figure 7 In the illustrated embodiment, when the third control signal c2 is not determined to have a value of 0, the third FF 734 is disabled, so the addition unit 738 outputs the previous addition result. However, the embodiments of the present disclosure are not limited thereto. For example, when it is determined that the third control signal (or the addition output holding signal) c2 has a value of 1, the third FF 734 is disabled.

[0070] Fig. 9 9 is a flow chart illustrating a process 900 performed by a neural network acceleration device (NNAD) according to an embodiment. In an embodiment, the NNAD includes a signal detector (e.g., Figure 4 signal detector 430 in the PE cluster (eg, Figure 3 The signal detector includes a signal converter (e.g., Figure 4 signal converter 412) and a flag generating circuit (e.g., Figure 4 The flag generation circuit 420 in the PE cluster includes a plurality of PEs (eg, Figure 7 750-1), each PE has a PE controller (e.g., Figure 7 PE controller 780) and computing units (e.g., Figure 7 The computing unit 790 in it).

[0071] In S910, a signal detector receives a first input signal (eg, Figure 4 The first input signal Feature) and the second input signal (for example, Figure 4 The second input signal Weight in .

[0072] In S930, the signal converter converts the first input signal and the second input signal into a first converted input signal (eg, Figure 4 The first converted input signal cf in the example) and the second converted input signal (eg, Figure 4 The second converted input signal cw in FIG. 2 is referred to as follows. Fig.10 The operation of the signal converter is described in more detail.

[0073] In S950, the flag generation circuit generates one or more flag signals in response to the first converted input signal and the second converted input signal, and provides the flag signals to the PE cluster. Fig.11 The operation of the flag generation circuit is described in more detail.

[0074] In S970, the PE controller controls the first internal memory array (eg, Figure 7 One or more read operations of multiple memories in the first internal memory array 760) and one or more operations of the operation unit. Fig.12 The operation of the PE controller is described in more detail.

[0075] Fig.10 is a diagram showing that a first input signal (eg, Figure 4 The first input signal Feature) and the second input signal (for example, Figure 4 The second input signal Weight in is respectively converted into the first converted input signal (for example, Figure 4 The first converted input signal cf in the example) and the second converted input signal (eg, Figure 4 Flowchart of process 1030 of the second converted input signal cw).

[0076] In S1020, a signal converter (eg, Figure 4 The signal converter 412 of FIG. 1030 determines whether the first input signal or the second input signal or both have a value of 0. If the signal converter determines that the first input signal or the second input signal or both have a value of 0, the process 1030 proceeds to S1090.

[0077] In S1050 , the signal converter determines whether the second input signal has a value of −1 or 1. If the signal converter determines that the second input signal has a value of −1 or 1, the process 1030 proceeds to S1060 ; otherwise, in S1050 , the process S1030 proceeds to S1070 .

[0078] In S1060, when the second input signal has a value of -1, the signal converter generates a first converted input signal that is an inverse sign version of the first input signal and a second converted input signal having a value of 1. When the second input signal has a value of 1, the signal converter generates a first converted input signal having the same value as the first input signal and a second converted input signal having a value of 1. Therefore, the first converted input signal has the same value as a multiplication result of the value of the first input signal and the value of the second input signal, and the second converted input signal has a value of 1.

[0079] In S1070 , the signal converter determines whether the first input signal has a value of −1 or 1. If the signal converter determines that the first input signal has a value of −1 or 1, the process 1030 proceeds to S1080 ; otherwise, in S1070 , the process 1030 proceeds to S1090 .

[0080] In S1080, when the first input signal has a value of -1, the signal converter generates a first converted input signal having a value of 1 and a second converted input signal which is an inverse sign version of the second input signal. When the first input signal has a value of 1, the signal converter generates a first converted input signal having a value of 1 and a second converted input signal having the same value as the second input signal. Therefore, the first converted input signal has a value of 1, and the second converted input signal has the same value as a multiplication result of the value of the first input signal and the value of the second input signal.

[0081] In S1090 , the signal converter outputs the value of the first input signal and the value of the second input signal as the value of the first converted input signal and the value of the second converted input signal, respectively.

[0082] In S1095 , the signal converter outputs the value of the first converted input signal and the value of the second converted input signal generated in S1060 or S1080 .

[0083] Fig.11 11 is a flow chart illustrating a process 1150 for generating one or more flag signals according to an embodiment. In an embodiment, the flag signals include a first zero-valued flag signal (e.g., Figure 4 The first zero value flag signal zff in the example), the second zero value flag signal (for example, Figure 4 The second zero value flag signal zfw in, the third zero value flag signal (for example, Figure 4 The third zero-value flag signal zf in the first one-value flag signal (for example, Figure 4 The first single-value flag signal off) and the second single-value flag signal ( Figure 4 The second one-value flag signal ofw in.

[0084] In S1110, a first flag generator (eg, Figure 4 The first flag generator 414 in the embodiment receives a first converted input signal (eg, Figure 4 The first converted input signal cf in the embodiment of the present invention is a first converted input signal cf in the embodiment of the present invention, and the second flag generator (eg, Figure 4 The second flag generator 416 in the embodiment receives the second converted input signal (eg, Figure 4 The first and second converted input signals may be Fig.10 The output of process 1030.

[0085] In S1130, the first flag generator generates a first zero-value flag signal and a first one-value flag signal in response to the first converted input signal. In an embodiment, when the first converted input signal has a value of 0, the first flag generator generates a first zero-value flag signal having a first value (e.g., 1) and a first one-value flag signal having a second value (e.g., 0); when the first converted input signal has a value of 1, a first zero-value flag signal having a second value and a first one-value flag signal having a first value are generated; otherwise, a first zero-value flag signal having a second value and a first one-value flag signal having a second value are generated. The second flag generator generates a second zero-value flag signal and a second one-value flag signal in response to the second converted input signal. In an embodiment, when the second converted input signal has a value of 0, the second flag generator generates a second zero-value flag signal having a first value (e.g., 1) and a second one-value flag signal having a second value (e.g., 0); when the second converted input signal has a value of 1, a second zero-value flag signal having a second value and a second one-value flag signal having a first value are generated; otherwise, a second zero-value flag signal having a second value and a second one-value flag signal having a second value are generated.

[0086] In S1150, the first flag generator sends a flag to a lookup table (eg, Figure 4 418 in the LUT 418) provides a first zero value flag signal and sends a signal to the PE cluster (e.g., Figure 3 The second flag generator provides a second zero-valued flag signal to the LUT and provides a second one-valued flag signal to the PE cluster.

[0087] In S1170, the LUT generates a third zero-value flag signal (or a final zero-value signal) in response to the first zero-value flag signal and the second zero-value flag signal, and provides the third zero-value flag signal to the PE cluster. In an embodiment, the third zero-value flag signal is the result of a logical OR operation on the first zero-value flag signal and the second zero-value flag signal.

[0088] Fig.12 is a diagram showing control of a first internal memory array (eg, Figure 7 One or more read operations of multiple memories in the first internal memory array 760) and an operation unit (e.g., Figure 7 Flowchart of process 1270 for one or more operations of the computing unit 790 in the process.

[0089] In S1210, PE (e.g., Figure 7 PE 750 in ) receives the converted input signal and the flag signal, and stores the converted input signal in a first internal memory array, and stores the flag signal in a second internal memory array (e.g., Figure 7 In an embodiment, the first internal memory includes a first high-speed scratchpad memory (e.g., Figure 7 The first high-speed scratch pad memory 760-1 and the second high-speed scratch pad memory (eg, Figure 7 The first high-speed temporary storage memory stores the first converted input signal (eg, Figure 7 The second high-speed temporary storage memory stores the value of the first converted input signal cf in the first input signal cf), and the second high-speed temporary storage memory stores the second converted input signal (for example, Figure 7 The second internal memory includes a first register array (eg, Figure 7 770-1), a second register array (eg, Figure 7 The second register array 770-2 in FIG. 770-1 and the third register array (eg, Figure 7 The first register array stores the final zero-value flag signal (e.g., Figure 7 The second register array stores the value of the third zero-value flag signal zf in the first register array, and the second register array stores the first one-value flag signal (for example, Figure 7 The value of the first one-valued flag signal off) in the third register array, and the third register array stores the second one-valued flag signal (for example, Figure 7 The value of the second one-value flag signal ofw) in.

[0090] In S1220, a PE controller (eg, Figure 7 The PE controller 780 in the first register array determines whether the value of the zero-value flag signal stored in the first register array is the first value (e.g., 1). When the PE controller determines that the value of the zero-value flag signal stored in the first register array is the first value, the process 1270 proceeds to S1230; otherwise, in S1220, the process 1270 proceeds to S1240.

[0091] In S1230, the PE controller controls to skip a first read operation on the first high-speed scratch pad memory and a second read operation on the second high-speed scratch pad memory. The PE controller causes the operation unit to skip a multiplication operation (eg, Figure 7 ) and addition operations (e.g., by Figure 7 The addition operation is performed by the addition unit 738 in .

[0092] In S1240, the PE controller determines whether the value of the first one-value flag signal or the value of the second one-value flag signal or both are the first value (e.g., 1). If the PE controller determines that the value of the first one-value flag signal or the value of the second one-value flag signal or both are the first value, process 1270 proceeds to S1250; otherwise, in S1240, process 1270 proceeds to S1260.

[0093] In S1250, the PE controller controls to skip the first read operation of the first high-speed scratch pad memory or the second read operation of the second high-speed scratch pad memory. The PE controller further enables the operation unit to skip the multiplication operation and perform the addition operation. For example, in an embodiment, when the first single-value flag signal is the first value, the PE controller skips reading the first high-speed scratch pad memory, reads the second high-speed scratch pad memory, and uses the value read from the second high-speed scratch pad memory to perform the addition operation; otherwise, the PE controller skips reading the second high-speed scratch pad memory; reads the first high-speed scratch pad memory, and uses the value read from the first high-speed scratch pad memory to perform the addition operation.

[0094] In S1260, the PE controller controls to perform a first read operation on the first scratch pad memory and a second read operation on the second scratch pad memory. The PE controller further enables the operation unit to perform a multiplication operation using the values ​​read from the first and second scratch pad memories and to perform an addition operation using the result of the multiplication operation.

[0095] Various aspects of the present disclosure have been described in conjunction with specific embodiments proposed as examples. Many substitutions, modifications and variations may be made to the embodiments described herein without departing from the scope of the appended claims. Therefore, the embodiments described herein are intended to be illustrative rather than restrictive.

Claims

1. An acceleration device, comprising: a signal detector that converts a first input signal and a second input signal into a first converted input signal and a second converted input signal, respectively, and determines a final zero-value flag signal indicating whether a value of at least one of the first converted input signal and the second converted input signal is 0, a first one-value flag signal indicating whether a value of the first converted input signal is 1, and a second one-value flag signal indicating whether a value of the second converted input signal is 1; as well as A processing element, i.e., PE, determines at least one arithmetic operation that does not need to be performed among multiple arithmetic operations performed by the PE based on the final zero-value flag signal, the first-value flag signal, and the second-value flag signal, and when the final zero-value flag signal has a first value, the PE skips the first arithmetic operation that is determined not to be performed and a second arithmetic operation different from the first arithmetic operation, and the first value of the final zero-value flag signal indicates that at least one of the first input signal and the second input signal has a value of 0.

2. The acceleration device according to claim 1, wherein when the first-value flag signal or the second-value flag signal or both have the first value and the final zero-value flag signal has a second value, the PE skips the first arithmetic operation and performs the second arithmetic operation, the first value of the first-value flag signal indicating that the first input signal has a value of -1 or 1, and the first value of the second-value flag signal indicating that the second input signal has a value of -1 or 1.

3. The acceleration device according to claim 1, wherein the PE comprises: a first internal memory array, the first internal memory comprising a first scratchpad memory and a second scratchpad memory, the first scratchpad memory storing the value of the first converted input signal and the second scratchpad memory storing the value of the second converted input signal, the first scratchpad memory and the second scratchpad memory being of a first type; a second internal memory array including a first register array storing a value of the final zero-value flag signal, the first register array being of a second type different from the first type; as well as a controller that reads the value of the final zero-value flag signal from the first register array, and skips a first read operation of reading the value of the first converted input signal of the first scratch pad memory and a second read operation of reading the value of the second converted input signal of the second scratch pad memory when the final zero-value flag signal has the first value.

4. The acceleration device according to claim 3, wherein the second internal memory array further comprises: a second register array storing the value of the first value flag signal, the second register array being of the second type; as well as a third register array storing the value of the second-value flag signal, the third register array being of the second type, wherein the controller reads the value of the first-value flag signal, or the value of the second-value flag signal, or both, and When the first-value flag signal or the second-value flag signal or both have the first value, and the final zero-value flag signal has a second value, the controller skips the first read operation or the second read operation.

5. The acceleration device according to claim 4, wherein the controller reads the value of the second-value flag signal from the third register array, and When the value of the second-value flag signal has the first value, the controller skips the second reading operation.

6. The acceleration device according to claim 4, wherein the controller generates a first control signal, a second control signal, and a third control signal based on the value of the final zero-value flag signal, the value of the first-value flag signal, and the value of the second-value flag signal, and Wherein the first arithmetic operation is a multiplication operation, and the second arithmetic operation is an addition operation, the PE further comprises: a multiplication circuit that performs a multiplication operation on a value of the first converted input signal and a value of the second converted input signal in response to the first control signal and the second control signal; as well as The adding circuit performs an addition operation on the multiplication result from the multiplication circuit and a previous addition result in response to the third control signal.

7. The acceleration device according to claim 6, wherein the multiplication circuit comprises: a first flip-flop, namely a first FF, coupled to the first scratch pad memory and enabled in response to the first control signal; a second FF coupled to the second high-speed scratchpad memory and enabled in response to the first control signal; a multiplier that performs the multiplication operation on an output signal from the first FF and an output signal from the second FF; as well as A multiplexer, ie, MUX, selects one of the value of the first converted input signal, the value of the second converted input signal, and the output result of the multiplier as the multiplication result in response to the second control signal, and provides the multiplication result to the adding circuit.

8. The acceleration device according to claim 7, wherein the adding circuit comprises: an adder, adding the multiplication result and the previous addition result to output a current addition result; as well as A third FF is coupled to the adder and is enabled in response to the third control signal.

9. The acceleration device of claim 1, wherein when the first input signal or the second input signal or both have a value of 0, the signal detector generates the first converted input signal having the same value as the first input signal and the second converted input signal having the same value as the second input signal.

10. The acceleration device of claim 9, wherein when the second input signal has a value of 1 or a value of -1 and the first input signal has a value other than 0, the signal detector generates the first converted input signal as an opposite sign version of the first input signal and the second converted input signal having a value of 1.

11. An acceleration device according to claim 10, wherein when the first input signal has a value of 1 or a value of -1 and the second input signal has a value other than 0, 1 and -1, the signal detector generates the second converted input signal as an opposite sign version of the second input signal, and generates the first converted input signal having a value of 1.

12. The acceleration device according to claim 1, wherein the signal detector comprises: a first flag generator that generates a first zero-valued flag signal and the first one-valued flag signal in response to the first converted input signal; a second flag generator, generating a second zero-valued flag signal and a second one-valued flag signal in response to the second converted input signal; as well as A lookup table is provided to generate the final zero-valued flag signal in response to the first zero-valued flag signal and the second zero-valued flag signal. 13 . The acceleration device according to claim 12 , wherein the final zero-value flag signal is a result of a logical OR operation of the first zero-value flag signal and the second zero-value flag signal.

14. A method for controlling an acceleration device, the method comprising: converting the first input signal and the second input signal into a first converted input signal and a second converted input signal, respectively; determining a final zero-valued flag signal indicating whether a value of at least one of the first converted input signal and the second converted input signal is 0, a first one-valued flag signal indicating whether a value of the first converted input signal is 1, and a second one-valued flag signal indicating whether a value of the second converted input signal is 1; determining at least one arithmetic operation that does not need to be performed among a plurality of arithmetic operations based on the final zero-value flag signal, the first-value flag signal, and the second-value flag signal; and When the final zero-value flag signal has a first value, a first arithmetic operation that does not need to be performed and a second arithmetic operation different from the first arithmetic operation are skipped, and the first value of the final zero-value flag signal indicates that at least one of the first input signal and the second input signal has a value of 0.

15. The method according to claim 14, further comprising: when the first-valued flag signal or the second-valued flag signal or both have the first value and the final zero-valued flag signal has the second value, skipping the first arithmetic operation and performing the second arithmetic operation, The first value of the first-value flag signal indicates that the first input signal has a value of -1 or 1, and the first value of the second-value flag signal indicates that the second input signal has a value of -1 or 1.

16. The method according to claim 14, further comprising: storing the value of the first converted input signal in a first scratchpad memory, the first scratchpad memory being of a first type; storing the value of the second converted input signal in a second scratch pad memory, the second scratch pad memory being of a first type; storing the value of the final zero-value flag signal in a first register array, the first register array being of a second type different from the first type; Read the value of the final zero-value flag signal from the first register array; and When the final zero-value flag signal has the first value, a first read operation of reading the value of the first converted input signal of the first scratch pad memory and a second read operation of reading the value of the second converted input signal of the second scratch pad memory are skipped.

17. The method according to claim 14, further comprising: When the first input signal or the second input signal or both have a value of 0, the first converted input signal having the same value as the first input signal and the second converted input signal having the same value as the second input signal are generated.

18. A processing system comprising: a memory storing data and outputting an input signal in response to a request signal from a host; An acceleration device receives the input signal and performs one or more given arithmetic operations on the input signal, wherein the input signal includes a first input signal and a second input signal, and the acceleration device includes: a signal detector that converts the first input signal and the second input signal into a first converted input signal and a second converted input signal, respectively, and determines a final zero-value flag signal indicating whether a value of at least one of the first converted input signal and the second converted input signal is 0, a first one-value flag signal indicating whether a value of the first converted input signal is 1, and a second one-value flag signal indicating whether a value of the second converted input signal is 1; as well as A processing element, i.e., PE, determines at least one arithmetic operation that does not need to be performed among multiple arithmetic operations performed by the PE based on the final zero-value flag signal, the first-value flag signal, and the second-value flag signal, and skips the first arithmetic operation that is determined not to be performed and a second arithmetic operation different from the first arithmetic operation when the final zero-value flag signal has a first value, and the first value of the final zero-value flag signal indicates that at least one of the first input signal and the second input signal has a value of 0.

19. The processing system of claim 18, wherein the PE skips the first arithmetic operation and performs the second arithmetic operation when the first-value flag signal or the second-value flag signal or both have the first value and the final zero-value flag signal has a second value, the first value of the first-value flag signal indicating that the first input signal has a value of -1 or 1, and the first value of the second-value flag signal indicating that the second input signal has a value of -1 or 1.