Splice-Core GEMM design cuts tile energy in INT4 inference workloads
An arXiv paper describes two processing element variants for matrix multiplication that report large energy reductions over a binary systolic array in 45nm post-layout results.
In plain words
Matrix multiplication is the core computation behind AI inference, and its hardware units dominate chip energy use. The Splice-Core paper proposes two processing element designs that split low-order and high-order bits for different treatment. In simulations at 45nm, the authors report roughly 77% lower tile energy than a binary systolic array across 26 INT4 workloads.
What was announced
The paper, posted on arXiv, presents Splice-Core, a temporal-unary/binary hybrid general matrix multiplication unit. It minimizes latency in low-order accumulation while keeping value-dependent execution for high-order activation bits. The authors describe two processing element variants: Splice-BP, which uses bit planes, and Splice-LUT, which selects a shared weight multiple in a single cycle.
In 45nm evaluations across 26 INT4 convolutional and large language model workloads, the authors report geometric-mean tile energy reductions of 77.4% for Splice-BP and 76.4% for Splice-LUT versus a binary systolic array. Splice-LUT is further credited with 44.4% higher geometric-mean tile throughput and a 6.13x lower geometric-mean energy-delay product. On evaluated W8A8 convolution and expert-projection workloads, energy fell 77.3-88.1%, and post-place-and-route INT4 area shrank 14% and 9% respectively.
What it does not show
- The evidence is the paper abstract only; full methodology, baseline configuration, and peer review status are not available in the supplied passages.
- All reported energy, throughput, EDP, and area figures are author-reported simulation and post-place-and-route results at 45nm, not independent measurements or fabricated silicon.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.