Graphics compilers trail nvcc by orders of magnitude on compute code quality
mike64_t · x · 2026-11-01
Developer mike64t compares GPU compiler ecosystems: code from graphics compilers and compute shaders is often orders of magnitude more suboptimal than what nvcc generates from comparable CUDA C, which in turn trails hand-tuned kernels. In a follow-up he notes that in compute workloads cubins are typically statically precompiled or built via ptxas in a JIT-like fashion, all in userspace; the libcuda userspace compat shim bridges driver versions but shouldn't affect device-side execution of identical cubins.
More from Infra
- Picked up a $1,100 RTX 3090 FTW3 for a headless local LLM server — seeking configs — NighthawkXL · 2026-11-01
- Sequoia's Shaun Maguire builds SpaceX model projecting $731B revenue by 2028 — FinanceYF5 · 2026-11-01
- The Great "Rack Spread": B300 GPU rents up 106% while token prices fall 58% — FinanceYF5 · 2026-11-01
- SK hynix sold out through 2029: AI selloff is about cash flow, not demand, argues Puru Saxena — danielrock · 2026-11-01
- Gemma 4 reportedly running in a space data center under Google's Project Suncatcher — Saboo_Shubham_ · 2026-11-01
- Do driver versions change GPU device-code fingerprints? Experts debate compilers vs cubins — matiasgoldberg · 2026-11-01