OpenAI's Models Rewrote AMD's GPU Circuits and Gained 14% in Five Months
OpenAI's models found up to 14% power, performance and area gains in AMD's flagship GPU design in under five months, and AMD has already sent some of those changes to manufacturing.
- OpenAI's models found up to 14% PPA gains on AMD's flagship GPU in under five months.
- AMD put some of the changes in a design already sent to manufacturing and independently verified single-digit CPU gains.
- AI-optimized circuits in Jalapeño, OpenAI's Broadcom-built inference chip, were proven on real silicon.
- In OpenAI's new RISC-V eval, GPT-6.2 Astra scores about twice GPT-6 Astra on PicoRV32.
- GPT-Synopsys, a model trained to run Synopsys EDA tools, will bring this to other chipmakers. No launch date yet.
- Caveat: the figures are self-reported, and OpenAI holds warrants for up to 160M AMD shares.
OpenAI says its models improved heavily optimized blocks in AMD’s flagship GPU by up to 14% in less than five months. The company equates that gain to roughly one year of conventional engineering progress. AMD included some model-generated changes in a recent design sent for manufacturing and added others to production design files for future versions.
Alongside the AMD results, OpenAI published details about model-assisted optimization of its Jalapeño inference chip, a RISC-V benchmark for measuring chip-design skills, and its partnership with electronic design automation vendor Synopsys. The results indicate that models can contribute to production circuit design when their output passes established verification and synthesis tools.
AMD moves model edits toward silicon
Power, performance and area, collectively known as PPA, are coupled design constraints. A smaller circuit can consume less power while missing its timing target, and a faster circuit can require more energy or silicon. Chip teams therefore evaluate candidate designs under fixed functional, timing and manufacturing requirements. OpenAI’s 14% figure is a block-level PPA measure, with no stated equivalent for whole-GPU performance.
| Target | Reported result | Validation or status |
|---|---|---|
| AMD flagship GPU IP | Up to 14% PPA improvement over a heavily optimized baseline | Some edits included in a design sent for manufacturing |
| AMD next-generation flagship CPU | Single-digit percentage PPA gains across key blocks | Model-generated designs verified by AMD |
| OpenAI Jalapeño | Improvements of 56% for a BF16 multiplier, 21% for an FP4 dot-product block and 10% for an FP32 accumulator; area reductions of 10% and 8% for matrix and SIMD units | Observed during silicon bring-up |
For the GPU work, OpenAI’s model rewrote circuit blocks and used Synopsys tools to check that each candidate preserved the original behavior. Candidates that passed verification were then evaluated for power, performance and area. This process filters out compact or fast designs that introduce functional errors.
Jalapeño puts the method on silicon
OpenAI began using GPT-4o to assist with Jalapeño in early 2025. Developed with Broadcom, the inference chip progressed from its first register-transfer level code, or RTL, to tape-out in about nine months. RTL describes a chip’s digital logic, while tape-out is the handoff of a completed design for fabrication.
OpenAI says fabricated Jalapeño chips reproduced the model-assisted optimizations during bring-up, the initial process of powering on and validating new silicon. Its published percentages apply to defined arithmetic and compute blocks. The company attributes the model’s role to block optimization and makes no claim that it generated Jalapeño’s overall architecture.
RISC-V tests optimization and greenfield RTL
To compare model generations, OpenAI created an evaluation based on two open-source 32-bit RISC-V processor cores: PicoRV32 and lowRISC’s Ibex documentation. Ibex has been fabricated in several chips, giving the benchmark a more established reference design.
- Optimize existing RTL. The model receives a working core and attempts to improve its implementation.
- Build from a specification. The model receives written requirements and generates a core from scratch.
The benchmark applies a hard validity gate before grading design quality. A candidate receives a score of zero if it fails a functional check or occupies more than 16 times the baseline core’s area. Its simplified scoring logic looks like this:
def score(design, baseline):
if not design.passes_simulation():
return 0.0
if design.area > 16 * baseline.area:
return 0.0
return ppa_score(design, baseline)Across 9,881 scored runs covering both cores, both tasks and multiple token budgets, the model identified as GPT-6.2 Astra produced the highest scores. At the roughly 112,000-token setting on PicoRV32, its score was about twice that of GPT-6 Astra. On Ibex, it averaged approximately 0.75 when generating a core from written requirements. That custom score supports comparisons within OpenAI’s evaluation and has no direct production-PPA interpretation.
All five tested models scored higher when generating a core from written requirements than when modifying existing RTL. OpenAI also identifies several gaps between the benchmark and production development, including limited feature requirements, a permissive area ceiling and the absence of post-synthesis timing checks.
Synopsys carries the work beyond AMD
Under a multiyear agreement, OpenAI and Synopsys are training GPT-Synopsys to operate the vendor’s electronic design automation tools. EDA software verifies RTL, synthesizes it into gates, places and routes those gates, checks timing and prepares a design for fabrication. Synopsys and Cadence supply much of this commercial toolchain.
The planned system will run an iterative workflow: invoke an EDA tool, inspect its reports, revise the design and repeat. Engineers will be able to specify goals such as PPA, verification coverage or timing closure, which confirms that signals arrive within the clock period. OpenAI and Synopsys say early technical work with customers is underway.
AMD’s deal links chip gains to OpenAI’s economics
OpenAI and AMD already have a substantial commercial relationship. OpenAI committed to 6 gigawatts of computing capacity across multiple generations of AMD GPUs and secured warrants for as many as 160 million AMD shares, equivalent to nearly a 10% stake under the announced terms. The warrants vest according to deployment volume and AMD’s share price.
Those terms give OpenAI two incentives to improve AMD hardware: better GPUs can reduce the cost of running its models, while successful deployments can increase the value of its potential equity position. The financial relationship also warrants caution when assessing performance claims issued by the partners.
The workflow shifts toward search and review
- AMD could shorten optimization cycles if similar gains extend across more blocks and future products.
- RTL and circuit engineers could use models to generate more candidate implementations, then select among designs that pass verification and synthesis.
- Verification teams would remain responsible for functional correctness, timing, physical constraints and signoff, potentially reviewing a larger volume of generated designs.
- Smaller chip companies could gain access to optimization techniques that currently require large specialist teams, depending on pricing and deployment requirements.
- Synopsys gains a model designed around its toolchain as Cadence develops competing agents for RTL, verification, implementation and signoff.
The evidence has firm boundaries
- Attribution: The five-month timeline and 14% GPU result come from OpenAI, which has commercial and financial interests tied to AMD.
- Scope: The 14% figure is a best-case result for unnamed GPU blocks. OpenAI has not published the block-level baselines, constraints or distribution of results.
- Independent checks: AMD verified the CPU work, while no independent audit of the GPU result or public design files has been released.
- Benchmark ownership: OpenAI created and ran the RISC-V evaluation. Its open-source cores are less complex than modern flagship CPU and GPU designs.
- Manufacturing status: AMD has sent a design containing some edits for fabrication, while Jalapeño provides the reported evidence from tested silicon. Foundry capacity remains a separate constraint.
Access and replication remain unresolved
GPT-Synopsys currently has no public product, API, price, deployment model or general release date. External teams will need enough access to reproduce the AMD-level gains, inspect the PPA objectives and constraints, protect sensitive design files, and run standard signoff flows. Independent results from production-scale blocks will establish how broadly the reported improvements transfer beyond OpenAI and AMD.