OpenAI models boost AMD flagship GPU PPA 14% in 5 months, matching senior architects
智东西 · wechat · 2026-11-19
OpenAI says its models raised the PPA of AMD's flagship GPU IP by up to 14% in under five months — roughly a year of human engineering work — with some changes already in AMD's taped-out and production designs. Its own inference chip Jalapeño saw 13% standard-cell area reduction in the vector unit and 10% in the matrix unit using model-proposed changes, validated during tape-out debugging; GPT-4o was already assisting chip design in early 2025.
Key points:
- Models can rewrite RTL for compute, control, and data-movement circuits, verified with Synopsys EDA tools; researchers say they reach senior microarchitect level
- On two open-source RISC-V cores (PicoRV32, Ibex), GPT-6.2Astra scored highest on throughput optimization; GPT-6.1Sol beat GPT-5.6Sol
- All five model generations scored higher designing cores from written specs than optimizing existing designs — partly because they can drop low-value features like performance counters
- Evaluation spanned 9,881 valid scoring runs, using IPC × Yosys-estimated frequency, with 0 points for failed checks or >16x area
Related event: OpenAI Models Deliver Up to 14% PPA Gains on AMD Flagship GPU IP(11 posts)→
More from Infra
- Oracle co-CEO admits 'classic mistake' selling data centers to towns — jeremyakahn · 2026-11-19
- Surface Laptop Ultra With NVIDIA RTX Spark Runs 94GB Model, Finishes 4+ Enterprise Agent Workflows Per Hour — ryanshrout · 2026-11-19
- Nace ships 71.8GB DeepSeek V4.1 Flash coding quant with 39% lower KL than rivals — m4rkGS · 2026-11-19
- Dev open-sources self-hosted model registry so inference workers boot when HF is down — cmm324 · 2026-11-19
- Architect Labs unveils Frostbite, an agent-co-designed CPU claiming 4.6x perf per watt over Arm — IgorCarron · 2026-11-19
- Nvidia's unusual delay in filing its 10-Q sparks Wall Street intrigue — analyst explains why — TechCrunch AI · 2026-11-19