Method for expanding read-write bandwidth of on-chip cache
By building multiple AXI-port storage matrices under the FPGA clock and adopting a new address algorithm, the read and write bandwidth of the HBM cache is expanded, the problem of insufficient bandwidth of the existing HBM is solved, and high-speed and multi-functional audio-visual processing is realized, reducing the System costs and space requirements.
Patent Information
- Application Number
- CN202411109110.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-12
AI Technical Summary
The existing HBM storage space is sufficient but the read and write bandwidth is insufficient to meet the complex processing needs, especially when multiple processing is required after 8K signals are written to the cache.
By generating timing control signals under the FPGA clock, a storage matrix containing M AXI-ports is constructed, and a processing mechanism with the same address bits and different data bits is adopted to write data equally into M AXI-ports, and when needed, from Multiple AXI-port readouts realize the extended read and write bandwidth of data.
High-speed, multi-functional audio and video processing is realized, reducing signal output delay and the number of external devices, and saving system construction costs and space.
Smart Images

Figure CN118945427A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of communication technology and broadcasting and television technology, and in particular to a method for extending on-chip cache read and write bandwidth, which is suitable for extending high-speed video and audio data read and write bandwidth on a chip. Background Art
[0002] The resolution of 8K signal is as high as 7680×4320, which is 4 times of today's 4K "UHD" and 16 times of 1080P "Full HD". In addition to resolution, 8K also has the characteristics of wide color bit width and high frame rate. The color depth is 12 bits and the frame rate is 120 frames per second (progressive scanning). The bandwidth of 8K signal is equivalent to 4 4K signals, which is about 12G Gbit / s×4=48Gbit / s. The IO bandwidth of one 8K signal needs double bandwidth, which is about 48Gbit / s×2=96Gbit / s.
[0003] As 8K signals become more and more popular, ultra-high-definition signal processing is usually completed using the cache particles that come with the high-performance chips in the device. The storage performance of cache particles is a key factor restricting high-performance computing at present. At present, high-bandwidth memory HBM (High Bandwidth Memory) based on TSV technology (silicon through-via technology) has become the first choice for high-performance computing. HBM is composed of multiple DRAM DIEs stacked into a 3D structure, using TSV technology to achieve signal sharing and distribution; the high number of I / Os brings high bit width, and each DRAM DIE of HBM contains multiple channels that can be accessed independently. Each channel contains multiple I / O ports with a bit width of 64 / 128 bits, making the total bit width of HBM up to 1024 bits.
[0004] HBM is mainly used in high-performance computing chips, and is mostly used to work with data center computing chips such as GPU / FPGA / ASIC. The read and write bandwidth of HBM's I / O (AXI-Port) is 256bit / s. High-performance chips usually provide a clock frequency of about 400MHz. Assuming that the frequency is 450MHz, the HBM read and write I / O is 256bit / s×450MHz=115.2Gbit / s. The I / O bandwidth provided by HBM is as follows: Figure 1 shown.
[0005] The existing HBM provides 115.2Gbit / s I / O bandwidth, and it takes 96Gbit / s to process the input and output of one 8K signal, which can meet the usage requirements. However, for complex processing requirements, such as when multiple processes need to be performed simultaneously after the 8K signal is written into the cache, the bandwidth cannot meet the requirements, and the read and write bandwidth needs to be increased again. Summary of the invention
[0006] In view of the problem that HBM storage space is sufficient but the read and write bandwidth for multiple demands is insufficient in existing practical use, the present invention provides a method for expanding the read and write bandwidth of on-chip cache, which can meet the needs of single devices for real-time multiple processing of indivisible ultra-high bandwidth and high-speed signals.
[0007] The present invention discloses a method for extending the read and write bandwidth of an on-chip cache, comprising:
[0008] When controlling access, a timing control signal for HBM is generated based on the FPGA clock;
[0009] Assign values to the AXI-ports. The corresponding address channels and HBM channels under each AXI-port have corresponding addresses to construct a storage matrix containing M AXI-ports.
[0010] When data is written, the data is evenly divided and written into the storage matrix from M AXI-ports; when data needs to be read, it is read from multiple AXI-ports at the same time.
[0011] As a further improvement of the present invention, when writing data to HBM through AXI-port, a new address algorithm with the same address bits and different data bits is adopted. The new address consists of a particle ID and a particle access address, and the algorithm controls the writing and reading of data to the new address.
[0012] Compared with the prior art, the present invention has the following beneficial effects:
[0013] The present invention provides an orderly write and read cache environment for high-speed, high-bandwidth audio and video signals by expanding the read and write bandwidth technology of HBM cache, so as to make full use of the functions of FPGA chips, realize a multifunctional and highly integrated audio and video processing stand-alone product, reduce signal output delay, reduce the number of external device access, and save system construction costs and system physical space. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a schematic diagram of the I / O bandwidth of the existing HBM;
[0015] Figure 2 A schematic diagram of the I / O bandwidth of HBM disclosed in an embodiment of the present invention. DETAILED DESCRIPTION
[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0017] The present invention is further described in detail below in conjunction with the accompanying drawings:
[0018] HBM's AXI protocol has five independent channels, namely: the read address channel (AR), which carries the address and control information necessary for the transfer transaction, including ARVALID, ARADDR, and ARREADY signals; the read data channel (R), which is the data to be read, including RVALID, RDATA, RREADY, and RRESP signals; the write address channel (AW), which carries the address and control information necessary for the transfer transaction, including ARVALID, ARADDR, and ARREADY signals; the write data channel (W), which is the data to be written, including WVALID, WDATA, WSTRB, and WREADY signals; and the write response channel (B), which provides a way for the device to respond to write transactions.
[0019] The present invention utilizes the characteristic that the address bits of HBM DRAM particles are the same. When writing data to HBM through the AXI port, a processing mechanism with the same address bits and different data bits is adopted to invent a new address algorithm. The new address is composed of a particle ID and a particle access address. The algorithm controls the writing and reading of data to the new address, and binds the number of particles according to the demand. Each piece of data is evenly written to the new address and read out when needed. The number of particles required is based on the bandwidth demand for reading. It is assumed that there are multiple processors reading data from the cache later. The required memory size is planned in advance, and the rest is left for other applications.
[0020] like Figure 2 As shown, the present invention divides the data of a signal into M parts, preferably equally divided into M parts, where M is a positive integer ≥ 2; and stores them on M particles, and the read and write bandwidths are expanded M times, thereby solving the problem of completing multiple processing operations on a chip for one ultra-high bandwidth, high-speed transmission signal, and achieving the result of integrating multiple functions on a single device.
[0021] Taking the 8K data, 450Ms / clock FPGA chip as an example, the data bandwidth requirement of 8K is 48Gbit / s, and the bandwidth provided by an AXI-port is 256bits×450M=115.2Gbit / s, which can meet the reading of 2 channels of 8K signals; if M AXI-ports work at the same time, M×15.2Gbit / s bandwidth can be achieved, which can adapt to the high-performance FPGA processing capabilities and give full play to the chip's capacity resources.
[0022] The specific implementation method of the present invention is:
[0023] When controlling access, a timing control signal for HBM is generated based on the FPGA clock, so that all AXI-ports work at the same timing.
[0024] Assign values to the AXI-ports. The corresponding address channels and HBM channels under each AXI-port have corresponding addresses, thereby constructing a storage matrix containing M AXI-ports.
[0025] When writing data, the data is evenly divided and written into the storage matrix from M AXI-ports; when it needs to be read, it is read from multiple AXI-ports at the same time; for example, when writing data, 8K audio and video data is evenly divided and written into the storage matrix from M AXI-ports according to the entire image; when it needs to be read, it is read from multiple AXI-ports at the same time, thereby realizing the expansion of data reading and writing bandwidth and meeting the needs of one input and multiple outputs.
[0026] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for extending the read and write bandwidth of on-chip cache, It is characterized in that include: When controlling access, a timing control signal for HBM is generated based on the FPGA clock; Assign values to the AXI-ports. The corresponding address channels and HBM channels under each AXI-port have corresponding addresses to construct a storage matrix containing M AXI-ports. When data is written, the data is evenly divided and written into the storage matrix from M AXI-ports; when data needs to be read, it is read from multiple AXI-ports at the same time.
2. The method for extending the on-chip cache read and write bandwidth as claimed in claim 1, It is characterized in that When writing data to HBM through AXI-port, a new address algorithm with the same address bits and different data bits is used. The new address consists of the particle ID and the particle access address. The algorithm controls the writing and reading of data to the new address.
