Invest in AI Hardware? OpenAI AMD GPU Gains Hit 14%

November 19, 2026
4 mins read
Close-up macro photograph of a bare integrated circuit die showing six gold bond pads and on-die interconnect traces.
A bare integrated circuit die with gold bond pads and interconnect traces, the kind of silicon OpenAI says its models helped redesign. (Epop / Wikimedia Commons, CC0 1.0)
If you buy, build, or invest in AI hardware, chip design speed now matters. On November 18, 2026, OpenAI said its AI models helped redesign key parts of an AMD GPU. A GPU, or graphics processing unit, is the chip type that does most of the heavy lifting for AI training and answering. Some of those changes are already heading to the factory. Here is what the results show, and what they do not yet prove. OpenAI published the results on November 18, 2026, in a technical post on its website. Its models rewrote parts of SystemVerilog code, the text that describes how circuits are wired inside a chip, according to OpenAI’s technical post. The target was AMD’s flagship GPU intellectual property (IP), a reusable block of chip design developed by AMD. The 14% figure is a best-case result for individual blocks of AMD’s flagship GPU IP, not a speedup of the whole GPU. Karmactive has also covered an AI method that cut wireless chip design time from weeks to minutes. OpenAI also reported that AMD verified single-digit percentage PPA gains across key blocks of its next-generation flagship CPU, the central processor that handles general computing tasks. Single-digit means below 10%, so the CPU improvements were smaller than the GPU result. The work builds on a commercial relationship. In October 2025, AMD and OpenAI announced a multi-year agreement to deploy 6 gigawatts of AMD Instinct GPUs, as AMD’s newsroom announced. The first 1-gigawatt deployment of AMD Instinct MI450 GPUs is set to begin in the second half of 2026. A gigawatt measures power capacity, roughly the output of a large power plant.

How OpenAI Redesigned Jalapeño Compute Blocks

OpenAI’s first custom chip, called Jalapeño, was developed with Broadcom and Celestica and is built for inference, the stage where a trained AI model produces answers. OpenAI says its models also redesigned compute blocks on that chip. A compute block is a group of circuits that performs one kind of math. The warp engine and the matrix unit are two of these blocks. The matrix unit handles the grid-style math at the heart of most AI models. After the redesign, the warp engine’s cell area fell by about 13%, and the matrix unit’s fell by about 10%. Cell area is the physical footprint of the basic logic building blocks on a chip. OpenAI says these optimizations were silicon-proven during Jalapeño’s bring-up, meaning they were confirmed on physical chips that had been built. Individual circuits showed larger savings. A multiplier in the matrix unit used about 15% less cell area in synthesis, and its latency dropped from two clock cycles to one. Latency is the delay before a result is ready. A clock cycle is one tick of the chip’s timing signal. A separate adder shrank by about 15% as well.

How OpenAI Tested Chip Design Skill

OpenAI also measured how well its models design processor logic from written specifications. The tests used two open-source RISC-V processor designs, called PicoRV32 and OpenTitan Ibex. RISC-V is a free, open standard for processor instructions, so anyone can build on it. In total, the company ran 9,881 scored runs. The runs covered several token budgets, and the headline comparisons used a comparable budget of about 112,000 tokens per run. A token is a small chunk of text a model reads and writes. The models compared were GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol, and GPT-5.5. On PicoRV32 at that budget, GPT-6.2 Astra scored roughly twice as high as GPT-6 Astra. For designs built from a written specification, GPT-6.2 Astra’s mean score was above 0.7 on PicoRV32 and about 0.75 on Ibex. Optimization was harder. When models built designs from a written specification, they could omit features found in the original cores, such as performance counters and additional instructions. Performance counters track processor or system events. Extra instructions are additional commands the processor can run. OpenAI says these omissions could explain part of the score advantage.

GPT-Synopsys and What Comes Next

OpenAI also described GPT-Synopsys, a training partnership to teach a model to work directly with Synopsys EDA software, confirmed in Synopsys’s own announcement. EDA stands for electronic design automation, the software engineers use to design, simulate, and check chips. The model is still in training. If it succeeds, an AI model would work inside the same software engineers use to design chips today. Engineers and investors will watch three things. First, whether timing rules, which make sure signals arrive on schedule, hold up in physical chip layout. Second, whether the AMD design with changes already sent to manufacturing performs as the figures suggest. Where advanced chips are made also matters, as Karmactive has covered in its report on TSMC’s $165 billion US chip investment. Third, how quickly chip teams adopt AI-written chip code. OpenAI reports that its models improved individual blocks of AMD’s flagship GPU IP by up to 14% on power, performance, and area in under five months. Some of those changes are already headed to manufacturing. On OpenAI’s Jalapeño chip, the warp engine and matrix unit shrank by about 13% and 10%, respectively. Benchmarks from 9,881 RISC-V test runs show GPT-6.2 Astra scoring about twice GPT-6 Astra on PicoRV32, though building from a specification still scored higher than optimizing existing code. Production silicon will be the real test of these results.

How OpenAI Improved AMD GPU Chip Design

An air traffic control tower against a blue sky
Previous Story

FAA Signed Off Arkansas Repair Minutes Before Nationwide Ground Stop

The chapel and quadrangle of Oriel College at the University of Oxford
Next Story

Hundreds Protest Nathan Cofnas Talk at Oriel College Over Jason Arday Row

Latest from Technology

Don't Miss