基于 RISC-V 的 FPGA GPGPU：一种具有竞争力的科学计算方法.pdf

1、RISC-V based GPGPU on FPGA:A Competitive Approach for Scientific Computing?Eric GuthmullerJrme FereyreRISC-V Summit Europe,2025-05-13 Scientific computing applications require 64b floating point computing precision Sometimes,64b is not even enough(see 1 E.Guthmuller,et al.,“Xvpfloat:RISC-V ISA Exten

2、sion for Variable Extended Precision Floating Point Computation”,(2024)IEEE Transactions on Computers)2025-05-13RISC-V Summit Europe 2025-Eric Guthmuller2Motivation2024-11 TOP500 GPGPUs have enabled exa-FLOPs class performance in recent supercomputers Codes have been adapted to GPGPU computing parad

3、igm(costly)But AI market is exploding and is much bigger than scientific computingGPGPUs are more and more optimized for low precision computingHow long before 64b support is dropped or emulated?Scientific computing applications require 64b floating point computing precision Sometimes,64b is not eve

4、n enough(see 1 Guthmuller E.,et al.,“Xvpfloat:RISC-V ISA Extension for Variable Extended Precision Floating Point Computation”,(2024)IEEE Transactions on Computers)2025-05-13RISC-V Summit Europe 2025-Eric Guthmuller3Motivation2024-11 TOP500 GPGPUs have enabled exa-FLOPs class performance in recent s

5、upercomputers Codes have been adapted to GPGPU computing paradigm(costly)But AI market is exploding and is much bigger than scientific computingGPGPUs are more and more optimized for low precision computingHow long before 64b support is dropped or emulated?Our objective:Explore the feasibility/perfo

6、rmance of a GPGPU implemented on FPGA with support for FP64 computation and targeting scientific computing use cases Scientific computing apps rely on well optimized linear algebra frameworks,e.g.BLAS Mostly vector-vector(lvl1)and matrix-vector(lvl2)RISC-V Summit Europe 2025-Eric Guthmuller4Brief in