Desktop supercomputing hardware platform based on DSP and FPGA multi-co-processing card

Abstract

The invention provides a desktop supercomputing hardware platform based on DSP and FPGA multi-co-processing card. The hardware platform is composed of several co-processing cards mounted on the computer chassis. The co-processing cards contain multi-DSP processors and multi-FPGA chips. The co-processing cards are expanded through the PCIE bus slot of the computer motherboard to construct a desktopsupercomputer hardware platform with high availability, low power consumption and small volume. A CPU and DSP-FPGA collaborative computing acceleration architecture is adopted in the invention, the number of DSP and FPGA chips on the co-processing card can be customized according to the task requirements, so that a single card can provide at least 1 trillion times per second of computing performance. The chassis supports the simultaneous expansion of up to 8 co-processing cards, beyond which the co-processing cards can be further expanded through the FMC daughter board, and the computing power provided by the whole chassis can reach several ten trillion times per second. In addition, the standard PCIE interface and FMC interface are used in the co-processing card, the high-openness DSP and FPGA are used in the computational architecture, and the general-purpose CPU is also used. Users can develop their own application programs according to the task requirements.

Description

technical field [0001] The invention relates to the technical field of desktop supercomputer hardware platforms, in particular to a desktop supercomputing hardware platform based on DSP and FPGA multi-coprocessing cards. Background technique [0002] In recent years, desktop supercomputers have broad application prospects in high-performance computing fields such as life sciences, engineering sciences, national defense technology, medical care, and finance. However, under the background of big data, the amount of data is increasing exponentially, and the complexity of task calculations is increasing. Higher demands are placed on desktop supercomputers. [0003] At present, the more common desktop supercomputing architecture is the CPU-GPU collaborative computing acceleration architecture, which uses CPU and GPU dual computing cores to achieve collaborative computing acceleration. Parallel computing tasks. This method has certain limitations: 1) The computing power is limit...

AI Technical Summary

Problems solved by technology

This method has certain limitations: 1) The computing power is limited, and the scalability is poor. The peak computing power of this architecture applied to the desktop supercomputing level is generally 4 trillion times per second, and it does not matter whether the GPU chip on a single processing node It is difficult to expand or expand GPU card processing nodes; 2) Poor openness, GPU-based desktop supercomputing is generally developed under the CUDA (Compute Unified Device Architecture, unified computing device architecture) architecture, and the CUDA architecture does not open the underlying driver interface, so that it can only use NVIDIA’s GPU chip as an accelerated computing card; 3) The power consumption of a GPU chip with a peak computing capacity of more than 1 trillion times per second is generally more than 200W. Cost, but also brings heat dissipation pressure and noise reduction pressure, and at the same time prevents GPU chips from being installed in high density, which limits scalability

Images