计算机科学
规范化(社会学)
Softmax函数
计算机硬件
人工智能
功能(生物学)
软件
计算机图形学(图像)
计算机视觉
硬件加速
作者
Shareefa Fairoose P.,Ashutosh Mishra
标识
DOI:10.1109/tvlsi.2026.3706101
摘要
Softmax computation in transformer-based LLM/SLM models commonly relies on online normalization to ensure numerical stability while reducing memory access cycles. However, conventional online formulations introduce loop-carried dependencies that result in serialization, restricting scalable spatial parallelism, and efficient wide-datapath realization in hardware implementations. This article presents a custom Softmax hardware accelerator (CSHA) based on an algorithm–architecture codesigned parallel online normalization framework. The proposed algorithm reformulates sequential online normalization into a parallel streaming-compatible structure while preserving numerical stability and memory efficiency, thereby enabling scalable datapath parallelism. All core computational units—maximum detection, exponential approximation, reduction trees, multiplication, and logarithmic computation—are circuit-optimized and synthesized in a 90 nm CMOS process to evaluate area, power, and delay tradeoffs. The CSHA employs a fully pipelined fixed-point architecture designed for accuracy-aware hardware efficiency. System-level validation is performed by replacing the floating-point Softmax in a pretrained BERT model with the proposed implementation, demonstrating less than 1% accuracy degradation across GLUE tasks and the SQuAD dataset. Field programmable gate array (FPGA) deployment on a Zynq-7000 SoC further verifies functional correctness and dynamic sequence-length configurability, eliminating redundant padded Softmax computation. By integrating algorithm reformulation with circuit-level optimization, the proposed CSHA provides an efficient and scalable oSftmax solution suitable for low-power, high-throughput inference systems.
科研通智能强力驱动
Strongly Powered by AbleSci AI