Kimi K3 case studies show kernel optimizations, a Triton-like compiler, and a chip prototype
teortaxesTex · x · 2026-07-28
The image shows Kimi K3 case studies covering three technical areas:
- GPU kernel optimization: On both NVIDIA Hopper and an alternative-vendor GPGPU, Kimi K3 improved AttnRes latency from 283.6 ms to 114.4 ms, cut DSA and KDA runtime by 55.1% and 73.6%, and reached over half of peak TFLOPS on MLA.
- GPU compiler development: Moonshot says Kimi K3 built MiniTriton, a Triton-like compiler with custom tile-level Python frontend, MLIR-based optimization, PTX codegen, and a PyTorch-like interface. On an NVIDIA L20, it outperformed PyTorch eager and torch.compile on its benchmark suite.
- Chip design: In a 48-hour autonomous run with Kimi Code, Kimi K3 reportedly built and verified an inference-chip prototype using open-source EDA tools and the Nangate45 standard-cell library. The design targets a 4 mm² budget, near 100 MHz, and more than 8,700 tokens/s simulated decode throughput.
The page links the compiler and chip work to Moonshot’s Kimi K3 research stack, with GitHub code for minitriton and nano-kpu.
More from Infra
- vLLM serves Beam at 1,769 tok/s on 4x B300 NVFP4, publishes Docker image and recipes — vllm_project · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Inside Beam: 8-of-256 expert routing across 52 layers, sliding window plus NoPE attention — vllm_project · 2026-10-16
- vLLM ships Day-0 support for Reflection-Beam-1.0, a 501B-param MoE with 1M context — vllm_project · 2026-10-16