Engineer questions the inline PTX hype DeepSeek sparked: hand-written PTX isn't a proxy for performance
mike64_t · x · 2026-09-14
Engineer mike64t questions the inline PTX obsession DeepSeek sparked, arguing that hand-written PTX isn't directly indicative of performance in the "just write it in assembly" sense. Ironically the claim holds truer for AMD, where there's essentially no way to reach roofline without inline asm blocks — largely blamed on poor LLVM register allocation. He adds he agrees agents hillclimbing a training stack are hard to compete with given competent human architecture guidance, but doubts that's what SpaceX is actually doing.
Related event: Engineer Questions Inline PTX Hype; Agent-Optimized Training Stacks Debated(2 posts)→
More from Infra
- Rumble's Neocloud Pivot Lands $13.7B Compute Deal With Anthropic — andersonbcdefg · 2026-09-14
- Running an LLM agent on a 512MB board with decoupled memory and live cross-machine migration — D777Castle · 2026-09-14
- Maia 200 hits ~12 TFLOP/s FP4 in 1mm²: density should be a first-class goal — thoefler · 2026-09-14
- Dev's 24/7 self-hosted AI stack: OpenWebUI, pidot, Tailscale, GLM and DeepSeek — andfanilo · 2026-09-14
- Dream Photonics' laser integration render caught mirroring the whole chip image — jwt0625 · 2026-09-14
- Qwen3.8-Flash-Next on M3 Ultra: 559 t/s prompt processing, 31 t/s generation — rm-rf-rm · 2026-09-14