DeepSeek V4.1 Flash writes CUDA kernels, tops KernelBench decode at 8.12% roofline
scaling01 · x · 2026-10-19
elliotarledge benchmarked DeepSeek V4.1 Flash (1017) on KernelBench-CUDA: all four of its hand-written CUDA kernels pass on an RTX Pro 6000, and it takes #1 of 23 on MegaQwen decode at 8.12% of roofline, edging out Claude Sonnet 5.5's 8.00%.
The interesting part is how:
- FP8 KV cache trick: the problem specifies bf16, but the model stores the decode KV cache in fp8 (e4m3, no scales), computes the error math in its trace, calls the 0.08 tolerance "real slack," and spends it
- The precision call holds up at cos 0.99994 vs reference at 32k context, so it counts under the rules — a deliberate accuracy-for-speed trade, not a free win
Other kernels: Fused MoE #8/23 (9.65% vs Sonnet 5.5's 11.30%), Native Sparse Attention #6/17 (0.266ms vs Sonnet's 0.072ms), Grid+MinGRU #16/20 (20.02%; the previous version had no valid kernel).
Related event: DeepSeek V4.1-Flash-1017 Tops Multiple Open-Weight Leaderboards(8 posts)→
More from coding & agent
- x402 agent-payment standard logs 75.4M transactions, $24.2M in 30 days — AiJohnAllen · 2026-10-19
- Dev runs 300 Claude Code agents in parallel with a verification layer — ctjlewis · 2026-10-19
- From BabyAGI to ChatGPT dots: 12 posts that map 4 years of personal agents — heyneighbor · 2026-10-19
- WebMCP directory indexes 38,000+ tools across 6,400+ sites for AI agents — AiJohnAllen · 2026-10-19
- Dev building agent-first GitHub alternative to kill manual merges to main — fforres · 2026-10-19
- A seven-layer knowledge graph stack for a law firm to align model behavior — SnooPeripherals5313 · 2026-10-19