Deep Dive — OpenAI's models rewrote AMD's GPU code for 14% PPA

Share
Deep Dive — OpenAI's models rewrote AMD's GPU code for 14% PPA

OpenAI published a technical account on November 18 saying its models found up to 14% better power, performance and area on AMD's flagship GPU intellectual property, measured against baselines AMD's engineers had already spent years optimizing, according to OpenAI. Some of those changes are already in a design AMD has sent for manufacturing. Others sit in its production design files for future versions.

On OpenAI's own chip the result is harder to argue with, because it was checked in silicon. Model-proposed changes cut standard-cell area by about 13% in Jalapeño's vector unit and 10% in its matrix unit, and the reductions survived the chip's bring-up — the moment a design stops being a file and starts being a part you can power on.

Those two facts, taken together, move AI-assisted chip design from a demo to a line item. Chip design is the last engineering discipline where being wrong is unrecoverable: a bad gate is a software bug you patch, and a bad block is a mask set you eat. If models can find a year of progress in five months inside a design that has already been optimized by the people who make it, the constraint on custom silicon is no longer engineering hours. It is verification capacity.

What the models actually changed

OpenAI says it started using GPT-o3-mini on Jalapeño in early 2025, and that even that model found non-trivial wins. The work is not floorplanning or place-and-route, the parts of the flow that have been automated for decades. It is register-transfer-level code — SystemVerilog, the hardware description language engineers write to specify what a circuit does — rewritten block by block to do the same thing in less silicon or fewer cycles.

The examples OpenAI gives are the kind of thing a senior microarchitect brags about in a design review. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one. A separate adder redesign took roughly 15% off cell area while holding its two-cycle latency. In the vector unit, the savings came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit, by exploring alternative logic implementations and by moving where the computation is divided across pipeline stages so hardware can be shared and intermediate storage dropped. In the matrix unit, gains accumulated across adders, multipliers and low-precision dot-product circuits.

Every rewrite was checked for behavioral equivalence with Synopsys tools, and the same tools were used to weigh the power, performance and area trade-offs. OpenAI's summary of the capability is that the model reasoned about RTL "at the level of an expert microarchitect."

That phrase is doing the heavy lifting, and it is worth separating what it claims from what it proves. The model is not inventing an architecture. It is proposing local rewrites inside a design and a verification flow that humans built, against a baseline the humans also built. What is new is that the proposals are good enough to survive equivalence checking and end up in a mask set.

What 14% buys, and what it doesn't

PPA is three numbers that trade against each other: how much power a block burns, how fast it runs, how much area it occupies. A single "up to 14%" figure does not tell you the split. OpenAI does not say whether that maximum came from a power win, a frequency win, an area win, or some mix, and it does not break the result down block by block. It also does not name the GPU. The disclosure is a maximum over the flagship GPU IP plus a claim that AMD verified single-digit percentage gains across key blocks of its upcoming flagship CPU.

The company's framing is that this represents roughly a year of progress in less than five months. That is an estimate, not a measurement: no baseline date, no definition of what a year of PPA progress normally looks like inside AMD, no accounting for how many candidate rewrites the model produced that were rejected. Treat it as a vendor's characterization of a vendor-run result, which is what it is.

What makes it more than marketing is adoption. A benchmark win is a chart. A change that AMD put into a design sent for manufacturing is a decision by an engineering organization with its own tape-out calendar and its own reputation on the line — one that has every incentive to reject a model's suggestion that does not hold up in synthesis and timing. OpenAI says AMD did exactly that, and added further changes to production design files for future versions. AMD has not published its own numbers or its own account of the collaboration, so the verification here is second-hand.

The evaluation is the part researchers will read

To show the capability is improving with each model generation rather than being one lucky prompt, OpenAI ran five model generations across two open-source 32-bit RISC-V cores, PicoRV32 and the Ibex core inside OpenTitan's security chip, on two tasks: improve an existing design's throughput, and design a core from a written specification.

The scoring rules are the interesting part. Each design was checked in simulation and scored zero if it failed a check or used more than 16 times the original core's area — a constraint that stops a model from "winning" by spending unlimited silicon. For designs that passed, OpenAI measured instructions per clock on workloads the models never saw, took the geometric mean, multiplied by the maximum clock frequency Yosys estimated, and normalized the best observed throughput on each core to 1.0. The full evaluation covered 9,881 scored runs.

At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, and on PicoRV32 its score was roughly twice GPT-6 Astra's. GPT-6.1 Sol also beat GPT-5.6 Sol on both cores. The generational ladder is clean, which is the minimum you would want before believing any of this generalizes.

Then the caveat, which OpenAI states itself: at that token budget, all five models scored higher designing a core from a written specification than improving an existing design. Part of that is that a specification lets the model omit features the original core had — performance counters, extra instructions — that cost area and do little for the test workloads. OpenAI says requiring those features, imposing a tighter area limit and checking timing after physical synthesis would make the test more representative of production design work. That is an unusually honest paragraph from a company with a chart to sell, and it is also the reason the RISC-V numbers should not be read as a prediction about a GPU.

Why this lands now

Custom silicon stopped being a hyperscaler hobby about three years ago. Google built TPUs behind its own fabric, Amazon built Trainium, Meta built MTIA, and Microsoft and OpenAI went to Broadcom for an accelerator designed around LLM inference rather than adapted from one. Jalapeño, OpenAI's first, went from initial design to tape-out in nine months and was announced in June as the first of a multi-generation platform, with Broadcom on silicon implementation and Celestica on boards and racks.

We covered that chip when Jalapeño's first benchmarks put OpenAI on Nvidia's turf: the interesting claim was never the per-watt number, it was the loop — a company using its own models to design the hardware that runs its own models. Thursday's post is that loop closing with numbers attached, and with a second company's silicon in the sample.

The comparison that matters is not Nvidia. A 14% block-level area win on an AMD GPU does not dent CUDA, and it does not change what a rack of merchant accelerators costs next quarter. The comparison is the design-services business. If a model can produce a year of PPA progress in five months, then the marginal cost of a second ASIC program falls, the amortization window on a custom chip shortens, and the case for buying somebody else's general-purpose part weakens. That is the same logic that pushed Samsung's chip lab onto Claude, where Samsung's chip lab bet on Claude: 15x speed, zero margin for error described verification work that normally takes a month finishing in two days.

The tools layer is already being rewired around it. OpenAI and Synopsys announced GPT-Synopsys on September 30, a multi-year agreement with revenue sharing in which a specialized model operates Synopsys EDA tools directly, runs on OpenAI-hosted infrastructure and plugs into Synopsys's own agent platform, Reuters reported at the time. Read the two announcements together and the shape is clear: the model that proposes the rewrite is being trained to drive the tool that checks the rewrite. The check is the moat — for Synopsys, and for whoever's model is allowed to hold the tools.

The skeptical case is inside the building

The strongest objection comes from OpenAI's own hardware chief. Richard Ho said last month that he still does not fully trust AI design changes, so every one is checked, The Next Web reported. That is not a hedge; it is the operating model. The model proposes, the flow disposes, an engineer signs off. Which means the throughput of this technique is set by how fast humans and formal tools can verify, not by how fast the model can generate candidates.

That has three consequences the post does not dwell on. First, the headline gain is a gain inside a verified subset: every one of those 14% improvements was one a verification flow accepted, and we have no visibility into how many plausible-looking rewrites died there or how much engineer time the screening consumed. Second, the result is unauditable from outside. "In production design files" is a claim only AMD can confirm, and AMD has not. Third, PPA wins measured at block level do not compose into a product-level win: a shipping GPU's performance per watt also depends on memory, packaging, clocks and how the software schedules work onto it, and none of that is in this report.

None of that makes the result small. It makes it a milestone with a measurement problem attached — the same problem every vendor-run capability claim has, and the reason the RISC-V evaluation, with its hidden workloads and its area cap, is more valuable to outside researchers than the AMD numbers are.

What to watch

The first real test is naming. AMD has not identified the GPU or the CPU blocks, and a block-level claim becomes a product claim only when the part ships and someone measures it. Watch the next flagship's performance-per-watt disclosures against AMD's own prior generation, and watch whether AMD's engineers say publicly what OpenAI says privately.

The second is whether independent groups reproduce the RISC-V results under the harder constraints OpenAI itself proposes — required features, tighter area, post-synthesis timing. That is a test anyone can run, and it would settle whether the generational ladder is a capability curve or a token-budget artifact.

The third is GPT-Synopsys. A model that operates EDA tools directly, sold with revenue sharing, sets a price for design intelligence. If it ships generally available and chip teams adopt it, the EDA seat becomes an API call and the design-services margin compresses from the other end. If it stalls in pilots, Thursday's numbers stay a very good demo from a company with a chip to finish.

The fourth is the response. Nvidia's answer to custom silicon has been to make custom silicon plug into its fabric, which is cheaper than competing with it. A 14% block-level PPA gain does not change that math by itself. A pattern of them, across two vendors and three chip generations, would.

If a model can find 14% in a design its owner spent years optimizing, what exactly are you paying a chip engineer for? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

AI agents wrote 135 research papers — 114 failed the honesty check

AI agents wrote 135 research papers — 114 failed the honesty check

Nine frontier agents, 15 research problems that nobody has answered yet, no time limit, and a real machine each. Not one of their 135 papers cleared the bar the scientists set — and 114 of them failed before quality was even discussed, because the paper claimed work that was never done or reported numbers that contradicted the agent's own logs. The benchmark is called Into the Unknown, built by a Carnegie Mellon-led team with researchers at US Department of Energy national laboratories…

Google's fix for agents that memorize their tests

Google's fix for agents that memorize their tests

Let an agent rewrite the scaffolding around its own model against a fixed set of test tasks and it learns the tests, not the job. Scores on the tasks it practiced climb; gains on anything new shrink or vanish. Google Cloud AI Research's answer, written up by The Decoder today, is not to lock the scaffolding down but to regulate how the search moves through it — and the numbers suggest the regulation, not the model, is what makes self-improvement stick. The catch in recursive…

NVIDIA's AVO agent clears every ARC-AGI-3 public level

NVIDIA's AVO agent clears every ARC-AGI-3 public level

NVIDIA says its long-horizon agent system, AVO, completed all 183 levels across the 25 environments of the ARC-AGI-3 public set with a perfect 100.00 score. The result is real and independently checkable against the official scorecard — but the more interesting claim is buried underneath it: the system lifted Claude Opus 5 from a roughly 30% model baseline to 100% by changing the harness, not the model. What actually happened ARC-AGI-3 is the latest installment of François Chollet's attempt…