PallasForge adds Hopper BF16 matmul and fused W8A16 kernel tutorials with CUPTI benchmarking
A_K_Nain · x · 2026-10-13
PallasForge has been updated with two kernel examples for Hopper GPUs, aimed at people learning to write Pallas kernels:
- Simple BF16 matmul starter: teaches tiling, safe input buffer reuse, and output tile ordering for better L2 reuse, with an accompanying animation.
- Fused W8A16 kernel: performs fused dequantization, converting INT8 weight tiles to BF16 in registers and feeding them straight to WGMMA without an extra shared-memory write. Notes explain how operand ordering affects register/shared-memory efficiency and the lifetime rules for converted weight registers.
The repo also shows proper kernel benchmarking with CUPTI to strip Python overhead that blockuntilready(...) misses on small kernels — the fused W8A16 kernel hit a median 21.09us, 1.95x faster than prepared BF16. For profiling, the guide recommends using XProf together with Nsight Compute, covering GPU Speed of Light, memory workload, warp stalls, and how to read the report numbers.
More from coding & agent
- FinOps for Agentic AI: taming skyrocketing cloud and token costs — DavidLinthicum · 2026-10-14
- QGIS bans all AI-generated code, and longtime users are walking away — MaxLenormand · 2026-10-14
- Gemini's Argon wrote 5,000 lines of Python to build its own ray-tracing engine — fofrAI · 2026-10-14
- Developer's AI agents built, critiqued and rebuilt a property-data site 9 times overnight — jamesbrooksco · 2026-10-14
- Open-Generative-AI hits 30k stars: 600+ free models as a Sora/Veo alternative — matchaman11 · 2026-10-14
- Hunting old Linux gadgets to add to your mesh so agents can run stuff on them remotely — _Stocko_ · 2026-10-14