FIGFLEX's processing element recodes integer weights in radix-4, fusing FP-INT multiply-accumulate into a unified partial-product reduction, with an INT3 optimization that reclaims the idle lane at odd precision.
Mixed-precision LLM accelerator turns lower bit weights into proportionally faster compute
A radix-4 bit-parallel datapath and GPTQ-aware quantizer let FIGFLEX match or beat GPU-level accuracy at 2 to 3 bit precision while raising throughput as precision falls.
Academic
Faraz Tahmasebi · Haocheng Xu · Andrew Trauger · Yomal Perera · Zhenyu Tang · Nael Chowdhury · +1 more
University of California, Irvine
Research Digest··2 min read
The authors present FIGFLEX, a co-designed accelerator and quantizer for LLM inference that keeps activations in FP16 and weights as low-bit integers, without dequantizing.
§