RISC-V Technical Session | Edge GenAI with Accelerated Softmax & GELU скачать в хорошем качестве

RISC-V Technical Session | Edge GenAI with Accelerated Softmax & GELU

While General Matrix-Matrix Multiplications (GEMMs) dominate the computational workload of Transformers, common non-linearities such as softmax and GELU can become significant bottlenecks, even with moderate GEMM acceleration. In this session, we present a Transformer acceleration template based on the PULP (Parallel Ultra Low Power) cluster, featuring 8 general-purpose RISC-V cores, a 24x8 systolic array GEMM accelerator based on the RedMulE architecture, and SoftEx, a novel accelerator for softmax and GELU nonlinearities. The RISC-V Technical Sessions provide in-depth discussions on the advancements across RISC-V International Committees, Special Interest Groups (SIGs), Task Groups, Horizontal Committees (HCs), and the wider RISC-V technical community.