OpenAI's GPT rewrites AMD flagship GPU RTL, gaining up to 14% PPA in 5 months
新智元 · wechat · 2026-11-22
OpenAI announced its GPT models rewrote the SystemVerilog RTL of AMD's highly optimized flagship GPU, lifting PPA (power/performance/area) by up to 14% in under 5 months — progress that took AMD engineers years. The AI-written circuits have been sent to fabrication.
Key points:
- Modifications span compute, control, and data-movement logic; every change was verified for behavioral equivalence with Synopsys EDA tools.
- OpenAI claims expert-level microarchitect RTL reasoning; Greg Brockman discussed the result at the DealBook summit and Lisa Su amplified it.
- The approach was first validated on OpenAI's own Jalapeño inference chip: a redesigned multiplier cut area 15% and went from 2 cycles to 1, verified in silicon.
- Benchmarks on PicoRV32 and Ibex RISC-V cores show GPT-6.2 Astra scoring highest per token budget; all five models scored higher designing from scratch than editing existing code.
- Next: co-training GPT-Synopsys to wield the full EDA toolchain like a top chip designer.
More from Infra
- An AI bubble fable: from free cash flow to private credit to annuity blowup — Svwang1 · 2026-11-22
- One RX 7900 XTX runs 524k context at 767 tok/s, 3x faster than llama.cpp — W61k3r · 2026-11-22
- How GPUs organize work: kernel launches, thread blocks and scheduling explained — blaizedsouza · 2026-11-22
- Alibaba's T-Head files for HK IPO at up to $80B, rivaling Baidu's Kunlunxin — 创业邦 · 2026-11-22
- Can llama.cpp RPC span a Mac mini, Mac Studio and dual RTX 3090s to run a 120B model? — ruffus_or · 2026-11-22
- LBNL report: data center load growth did not drive up residential electricity prices (2019-2025) — Eliv_nurotic · 2026-11-22