
OpenAI Chip Design Results: Up to 14% PPA Gains on AMD GPU
OpenAI says its models delivered up to 14% better power, performance and area on AMD's flagship GPU design, and AMD has sent some changes to the fab.
OpenAI chip design work has moved from an internal experiment to shipping silicon at a major vendor. In a research post dated November 18, OpenAI reported that its models found up to 14% improvement in power, performance and area (PPA) on AMD's flagship GPU design, measured against baselines that human engineers had already tuned heavily. AMD has included some of those changes in a design it recently sent for manufacturing, according to the post.
- AMD GPU result: up to 14% PPA improvement on flagship GPU IP, achieved by rewriting SystemVerilog RTL.
- Speed of progress: OpenAI describes it as roughly a year of gains reached in under five months.
- AMD CPU result: single-digit percentage PPA gains across key blocks of AMD's next-generation flagship CPU, verified by AMD.
- In-house proof: on OpenAI's own Jalapeño inference chip, model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit.
What Did OpenAI's Models Actually Change?
Chips are described in a hardware language, here SystemVerilog RTL, before they become transistors. OpenAI says its model rewrote circuits that perform calculations, control operations and move data, then used Synopsys tools to confirm each rewrite still behaved identically and to measure the trade-offs in power, speed and area. The company characterizes the reasoning as on par with an expert microarchitect.
The Jalapeño examples make this concrete. One multiplier redesign used about 15% less cell area while dropping latency from two clock cycles to one. A separate adder redesign saved a similar amount of area at the same latency. OpenAI notes these optimizations were proven in silicon during the chip's bring-up, which matters because many AI-for-hardware claims stop at simulation.
These are OpenAI's own figures, so treat the comparative framing as the vendor's claim. The AMD numbers carry more weight because AMD verified the CPU gains and committed GPU changes to manufacturing. Tom's Hardware reported the same figures on November 20.
How Fast Is AI Chip Design Improving?
To track progress between model generations, OpenAI tested five models on two open-source 32-bit RISC-V cores, PicoRV32 and Ibex, across 9,881 scored runs. With a budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average throughput score on both cores, roughly double GPT-6 Astra on PicoRV32.
One finding is counterintuitive. Every model scored higher designing a core from a written specification than improving the existing design. OpenAI is candid about why that may flatter the results: starting fresh lets a model omit features such as performance counters that do little for the test workloads. The post says tighter area limits and post-synthesis timing checks would make the evaluation more representative of production work.
Why AI-Assisted Chip Design Matters Beyond AMD
A 14% PPA gain is the kind of step that normally comes from a process node change or a full redesign cycle. Getting it from software review of existing RTL suggests a compounding loop, where better models help build the chips that train and run better models. It also fits a wider shift we covered when agentic chip design tools attracted major funding earlier this year.
Next, OpenAI says it is partnering with Synopsys to train a dedicated model, GPT-Synopsys, to operate electronic design automation tools the way expert designers do. For readers following artificial intelligence research, the takeaway is that frontier models are now contributing verified improvements to production hardware, not only writing software.
Sources: OpenAI — November 18, 2026; Tom's Hardware — November 20, 2026.
More AI Stories

Gemini API Tool Retrieval: Load Tool Schemas Only on Demand
Gemini API Tool Retrieval lets agents keep hundreds of tools on standby and load schemas only when needed, cutting input tokens at no separate charge.

Qwen 4 Max Explained: 2.4T Parameters, $1.50 API Pricing
Alibaba's Qwen 4 Max pairs 2.4 trillion parameters with a 1M-token context at $1.50 per million input tokens. Here is what developers get today.

Gemini 4 Neon: Carbon-Class Smarts at $3 Per Million Tokens
Gemini 4 Neon scores 61 on the Artificial Analysis index at $3 per million input tokens and cuts cost per task 32% versus Carbon. See who should switch.
