OpenAI Models Deliver Up to 14% PPA Gains on AMD Flagship GPUs
OpenAI says its models delivered up to 14% PPA gains on AMD flagship GPU IP, cut 13% of standard-cell area in its Jalapeño chip, and will pair with Synopsys EDA tools via GPT-Synopsys.
By Amara Osei
3 min read

What's News
- OpenAI reported up to 14% PPA gains on AMD flagship GPU IP on November 18, 2026, with some rewrites headed to manufacturing.
- Model changes cut standard-cell area by ~13% in Jalapeño's warp engine and ~10% in its matrix unit; tapeout took nine months.
- OpenAI scored 9,881 runs across five models on two RISC-V cores; GPT-6.2 Astra led both at ~112,000 tokens per run.
- OpenAI partnered with Synopsys on September 30, 2026 to train GPT-Synopsys on EDA workflows with revenue sharing.
OpenAI reported on November 18, 2026 that its models achieved up to 14% improvement in power, performance and area (PPA) over heavily optimized baselines on AMD's flagship GPU intellectual property — and some of those circuit rewrites are already bound for manufacturing.
The gains came from rewriting SystemVerilog RTL, the code describing a chip's circuits. The model used Synopsys tools to confirm each rewrite preserved a block's behavior and to weigh PPA trade-offs. AMD folded additional model-proposed changes into production design files for future chip versions. OpenAI said the GPU work compressed roughly a year of progress into under five months, and AMD verified single-digit percentage PPA gains across key blocks of an upcoming next-generation flagship CPU built with model-written designs.
What did the models change in Jalapeño?
OpenAI started using GPT-4o in early 2025 to help design Jalapeño, its first custom inference chip. The team estimates model-proposed changes cut standard-cell area by roughly 13% in the warp engine — the unit managing parallel computing tasks — and by 10% in the matrix unit, an arithmetic component. The models explored alternative logic implementations, split computations across pipeline stages, shared hardware and trimmed intermediate storage.
The specifics are concrete. A matrix-unit multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one. A separate adder redesign trimmed cell area by about 15% and kept two-cycle latency. Dot-product logic and pipeline improvements added further savings, and the optimizations were proven in silicon during Jalapeño's bring-up. OpenAI's August 25, 2026 silicon results said AI assistance took Jalapeño from initial design to tapeout in nine months.
How did five model generations compare?
OpenAI benchmarked GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol and GPT-5.5 on two open-source 32-bit RISC-V cores: PicoRV32, a compact core built for FPGAs, and Ibex, the embedded processor inside OpenTitan's security chip. The evaluation spanned RTL optimization and building a core from written requirements and interface templates — 9,881 scored runs across tasks, cores and token budgets.
In optimization, each model got the original hardware description and a mandate to raise throughput. A design scored zero if it failed simulation or exceeded 16 times the original core's area. Passing designs earned a geometric mean of instructions per clock across hidden workloads, multiplied by Yosys-estimated maximum clock frequency; the best observed throughput per core scored 1.0, and averages counted the zeros.
At roughly 112,000 tokens per run, GPT-6.2 Astra posted the highest average on both cores in RTL optimization, and GPT-6.1 Sol beat GPT-5.6 Sol on both. In specification-to-RTL design, GPT-6.2 Astra scored above 0.7 on PicoRV32 and averaged about 0.75 on Ibex. All five models scored better designing from requirements than optimizing existing RTL.
Why might spec-based scores overstate capability?
OpenAI flagged a caveat. Models starting from written requirements could omit features present in the original cores — performance counters, extra instructions — that did little for throughput on test workloads. Those omissions may explain part of the advantage. Requiring the features, tightening area limits and checking timing after physical synthesis would make the test more representative of production design.
What comes next?
OpenAI is training GPT-Synopsys with Synopsys to use electronic design automation tools the way expert chip designers do, building on a multi-year agreement announced September 30, 2026 covering joint development, go-to-market work and revenue sharing. Synopsys said engineers would delegate objectives from PPA optimization to timing and verification closure, with agents running tools, interpreting results and iterating toward verified outcomes for engineer review. Synopsys added that customer design data won't train GPT-Synopsys and will stay encrypted at rest and in transit, with configurable retention, audit and permission controls. The open question — joining model-driven design gains to trusted verification — is exactly what GPT-Synopsys is built to answer.
Source: unite.ai
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,336 articles