Splice-Core: Low-Order-Bit-Splicing GEMM for Energy-Efficient AI Accelerators
Splice-Core: Low-Order-Bit-Splicing GEMM for Energy-Efficient AI Accelerators
Splice-Core is a temporal-unary/binary (tub) hybrid GEMM unit that minimizes low-order accumulation latency while retaining value-dependent execution for high-order activation bits, with two PE variants: Splice-BP using bit planes and Splice-LUT selecting a shared weight multiple in a single cycle. Compared to a binary systolic array in 45nm across 26 INT4 CNN and LLM workloads, Splice-BP/Splice-LUT achieve 77.4%/76.4% geometric-mean tile energy reductions, and Splice-LUT further raises geometric-mean tile throughput by 44.4% and reduces geometric-mean EDP by 6.13x. Across evaluated W8A8 convolution and expert-projection workloads, Splice-BP and Splice-LUT reduce tile energy by 77.3-88.1%, and post-place-and-route INT4 yields 14%/9% smaller area for Splice-BP/Splice-LUT versus the systolic array.
Source: Hacker News AI · arxiv.org