arxiv.org/abs/2610.30480 FIGFLEX: A Mixed-Precision FP-INT GEMM Accelerator Co-Designed with Low-Bit LLM Quantization
FIGFLEX is an algorithm-hardware co-designed accelerator for mixed-precision FP-INT matrix multiplication in weight-quantized LLMs, paired with a new quantizer. On LLaMA-2-7B it reports WikiText-2 perplexity of 5.85 (3-bit) and 8.77 (2-bit), beating GPTQ, AWQ and others, and 1.56x and 1.42x higher throughput per area than AxCore and FIGLUT.
— via arXiv Computer Architecture, Faraz Tahmasebi, Haocheng Xu, Andrew Trauger, Yomal Perera, Zhenyu Tang, Nael Chowdhury, Sitao Huang
