FlashPrefill V2 reports a 47x context speedup

FlashPrefill V2 reports up to a 47.26× speedup over FlashAttention-2 for 128K-token prefilling in FP8 on Nvidia H20 GPUs, with a 27.19× gain in BF16, according to the arXiv paper. The kernel adds mean-error correction, paged KV cache, continuous batching, and FlashAttention-3/4-aligned memory work, but the figures are the authors’ benchmark—not a production guarantee. Inference teams should profile prefill separately from decode before buying more hardware, alongside the archive’s measured economics of compressed local models.