SystolicAttention: Fusing FlashAttention within a Single Systolic Array

Abstract: Transformer models rely heavily on scaled dot-product attention (SDPA), typically implemented using the FlashAttention algorithm. However, current systolic-array-based accelerators face significant challenges when executing FlashAttention. Systolic arrays can only achieve high utilization for consecutive and large matrix multiplications. In contrast, FlashAttention requires frequently interleaved matrix multiplications and softmax operations. The frequent data swaps between the systolic array and external vector units result in low systolic array utilization. This is further exacerbated by the fact that softmax involves numerous non-matrix operations, which are not well-suited for systolic arrays. Moreover, the concurrent execution of matrix multiplication on systolic arrays and softmax on vector units leads to register file and SRAM port contention, further degrading performance. To overcome these limitations, we propose FSA, an enhanced systolic array architecture that enables the entire FlashAttention algorithm to run entirely within a single systolic array, eliminating the need for external vector units. At the core of FSA is SystolicAttention, a novel scheduling algorithm that maps FlashAttention operations onto systolic arrays with fine-grained, element-wise overlap. This significantly improves array utilization while preserving the original floating-point operation order to maintain numerical stability. We implement FSA in synthesizable RTL and evaluate its performance against state-of-the-art commercial accelerators. Our results show that FSA achieves 1.77x and 4.83x higher attention FLOPs/s utilization compared to AWS NeuronCore-v2 and Google TPUv5e, respectively, with only about 10% area overhead.