Kivorent News

Thursday, November 19, 2026

Tech and AINews

OpenAI says its models rewrote AMD GPU circuits for up to 14% gains

OpenAI says its models rewrote parts of AMD's flagship GPU design for up to 14% better power, performance and area, and that AMD sent some of the changes to manufacturing.

Kiln, covering chips, data centers and power
, 3 min read

14%Best PPA gain OpenAI reports on AMD's flagship GPU IP

What to know

  • OpenAI said on November 18 that its models found up to 14% power, performance and area (PPA) improvements on AMD's flagship GPU IP, measured against heavily optimized baselines.
  • OpenAI says AMD put some of the changes into a design it recently sent for manufacturing and verified single-digit PPA gains in blocks of its next-generation flagship CPU.
  • On OpenAI's own Jalapeño inference chip, model-written changes cut estimated standard-cell area by about 13% in the warp engine and 10% in the matrix unit.

The most telling line in OpenAI's chip-design post isn't the 14%. It's that AMD has already put some of the model's circuit rewrites into a design it sent to the factory. Benchmarks on open-source cores are easy to wave away. A design sent for manufacturing is a commitment no chipmaker makes lightly.

What OpenAI claims

In a research post on November 18, OpenAI said its models rewrote the SystemVerilog RTL, the code that describes a chip's circuits, for blocks of AMD's "flagship GPU IP." The best result was a 14% improvement in power, performance and area (PPA) against what OpenAI calls heavily optimized baselines. OpenAI describes that as roughly a year of progress in under five months. Twelve months of gains in five is a pace about 2.4 times faster than usual, if the comparison holds.

OpenAI also says AMD verified single-digit PPA gains across key blocks of its upcoming flagship CPU, using designs the model wrote. Some changes went into the design AMD recently sent for manufacturing; others were added to production design files for later versions.

The model checked its own work with Synopsys tools, which simulate whether each rewrite keeps the block's behavior and estimate the trade-offs. OpenAI is also training a model with Synopsys, called GPT-Synopsys, to drive those design tools directly.

The gaps

Three things are missing. AMD hasn't said anything in its own name; every AMD result in the post comes through OpenAI, and the two independent reports we read disagree on whether AMD itself confirmed the CPU figure. No chip is named. Analytics India Magazine notes the companies' 2025 deal started with MI450 GPUs, and we reported this week that TensorWave will run 457 MW of MI455X capacity for OpenAI, but nobody ties this week's work to either part. And "up to 14% PPA" is one figure for three measures. OpenAI gives no split, so a reader can't tell whether the gain is in watts, speed or silicon area.

Implicator.ai makes the fair point that tool-level checks don't establish how a finished product performs. That matters most for power, the number data center operators pay for, and only measured silicon will settle it.

Its own chip first

OpenAI has more to show on its own hardware. It says it began using GPT-4o on Jalapeño, its in-house inference chip, in early 2025. Model-written changes cut estimated standard-cell area by about 13% in the warp engine and 10% in the matrix unit, and one multiplier redesign used about 15% less area while cutting latency from two clock cycles to one. OpenAI says those were proven during Jalapeño's silicon bring-up, which is the strongest claim in the post because it rests on a chip that exists.

The public benchmark is narrower: two small open-source RISC-V cores, PicoRV32 and Ibex, tested in simulation. GPT-6.2 Astra, released earlier this month, scored highest.

For a beat that mostly counts chips by the megawatt, this is the other side of the ledger: whether models can squeeze more work out of each one. It ties into OpenAI's claims that agents now speed up its own research. If you want the background on why GPU blocks are laid out this way, start with our GPU vs CPU explainer. The question to watch is simple: when AMD's next flagship ships, does AMD say which parts a model drew?

Sources

AI chips and data centersOpenAIAI models

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.