Hierarchos-Native v0.2.5 ships P3 FP32-merge for RDNA3 WMMA GEMM, up to 27% faster
PhysicsDisastrous462 · reddit · 2026-11-22
Hierarchos-Native v0.2.5 introduces P3: the eight accumulator-bank partials in the RDNA3 full-FP16 WMMA kernel are now summed in FP32 with a single final FP16 rounding, preserving the existing two-wave workgroup structure.
- Numerics: 37/37 gates passed, 32/32 preserved K=96 outputs bit-identical, REF64F16 max error improved from 0.03035 to 0.02594, 31/32 cases improved with none worse
- Performance: four protected workloads measured at 1.02x–1.27x (max 27%), per-dispatch GEMM only, not end-to-end training speedups
- Also: per-batch profitability admission (16M MAC floor unchanged), a portable statically-linked Windows x64 package with 7 executables requiring only a Vulkan driver; WMMA stays opt-in (select full-fp16); qualified only on RDNA3 Phoenix (PCI 1002:15bf)
Related event: Hierarchos-Native v0.2.5 Boosts RDNA3 FP16 GEMM by 27%(2 posts)→
More from Infra
- WSJ: Associate of Jensen Huang's Partner Accused of Funneling Nvidia Chips to China — GaryMarcus · 2026-11-22
- prime-rl 0.11.0 ships: 1.5x faster training, new pretraining entrypoint, Mega MoE dispatcher — samsja19 · 2026-11-22
- Math suggests 7.7M GPUs would serve everyone: fuel for the AI overbuild hypothesis — bookwormengr · 2026-11-22
- Chinese NAND maker YMTC promotes ZHITAI SSDs at Korea's G-Star game show — terryyuezhuo · 2026-11-22
- SGLang runs Inkling 2 Mini in 65GB on DGX Spark, 2x faster prefill than vLLM — AIFlow_ML · 2026-11-22
- Open-source Agent Skill hill-climbs local LLM serving speed, users report up to 5x gains — CatAstro_Piyush · 2026-11-22