Anthropic says Claude ported itself to AMD GPUs in 3 days, writing ~100% of kernels unassisted
AccBalanced · x · 2026-11-11
Anthropic's roadshow video reveals that when the company adopted AMD GPUs for the first time this year, what used to take dozens of engineers months took just 3 days with essentially no human intervention — Claude ported or wrote almost 100% of the kernels.
The retweeted essay argues this presages CUDA's erosion: like hand-tuned 1980s assembly (PKZIP, VisiCalc) losing to C and optimizing compilers, AI training/inference code will favor automated compilation over hand-written kernels — eliminating the "CUDA tax" behind Nvidia's 90%+ margins. CUDA may become a high-level specification language à la Verilog: developers describe massively parallel algorithms in CUDA, then feed the source to an LLM that ports it to Cerebras, Amazon Trainium2, or Google TPUv6 low-level code — likely already within reach of the latest models, and generally feasible within a year or two.
More from coding & agent
- One prompt, 2.5 hours: Fable 5.2 builds a full Bugatti Tourbillon in Blender — FinanceYF5 · 2026-11-11
- Molted launches AI agent hosting from $2.88/month with browser, memory and 1,500+ integrations — nikola_mr64990 · 2026-11-11
- OpenAI's Dot Agent Emails a Fireplace Seller and Negotiates $1,000 Off Overnight — dkundel · 2026-11-11
- REA: Open-source MCP that turns AI agents into a full reverse-engineering studio — nikola_mr64990 · 2026-11-11
- Screenshot app into tldraw via MCP, annotate on iPad, let Claude iterate — floguo · 2026-11-11
- Open-source Claude plugin auto-generates full music videos from a single song — AIandDesign · 2026-11-11