US-based OpenAI reports up to roughly 14% PPA improvement in GPU design work with US-based AMD

Text
iHOLYSHITVISUAL / TECHChipsiHOLYSHIT / ILLUSTRATION
iHolyShit illustration

On November 18, US-based OpenAI said model-assisted circuit-code rewrites with US-based AMD delivered up to roughly 14% improvement in power, performance and area metrics for flagship GPU design blocks against highly optimized baselines. That does not mean a graphics card benchmarks 14% faster. OpenAI says some changes entered a design sent for manufacturing. Its open-source processor evaluations also show potential, but preserving features, tightening area limits and checking physical timing remain important steps toward production designs.

US-based OpenAI said on November 18 in its official blog, “Advancing chip design with OpenAI models,” that work with US-based AMD achieved up to roughly 14% improvement in PPA for flagship GPU IP against highly optimized baselines through model-assisted circuit-code rewrites. A GPU is a graphics processor, and IP here means reusable design blocks, not a graphics card already on sale. Its official account posted the news at 17:05:52 UTC on November 18 (1:05:52 a.m. on November 19, Beijing time).

PPA covers three measures considered together in chip design: power, performance and area. The figure of up to roughly 14% describes improvement in design metrics relative to a baseline, not a graphics card running benchmarks 14% faster. For ordinary users, the significance concerns how future chips might use electricity and silicon space more efficiently, rather than an immediate equivalent speed boost for cards currently available.

OpenAI says the work delivered roughly a year of progress in less than five months, describing that pace as unprecedented at the frontier of chip design. According to OpenAI, AMD incorporated some changes into a recent design sent for manufacturing, while other changes entered production design files for future revisions. Sending a design for manufacturing is a different stage from consumers being able to buy GPUs incorporating those changes, and does not mean complete graphics cards are already in mass production.

The method centers on models reasoning about and rewriting SystemVerilog RTL: code describing how circuits connect and operate. The changes affect computation, control and data movement. Tools from US-based Synopsys then verify that a block’s behavior remains unchanged and assess trade-offs among power, performance and area. The role is to assist design and verification, not to take over all engineering work or bypass human sign-off.

The approach has a background in in-house chip development. OpenAI’s first custom inference chip, Jalapeño, was announced in August this year. Its development used frontier models for acceleration, with GPT-4o assistance beginning in early 2025. The new blog revisits that work before discussing the transfer of the method to AMD’s third-party hardware. OpenAI also says relevant optimizations became “silicon proven”—validated on real silicon—during “silicon bring-up,” when manufactured chips are first powered on and tested.

The changes operate at a fine-grained circuit level. In the vector unit, which manages and executes parallel computing tasks, OpenAI attributes savings to floating-point adders and multipliers, conversion, reduction and special-function units. Models explored different logic implementations and pipeline divisions, as well as shared hardware and reduced intermediate storage.

For example, OpenAI says one multiplier in the matrix unit had roughly 15% less standard-cell area after synthesis, with latency falling from two cycles to one. An adder used roughly 15% less area while retaining two-cycle latency. Standard cells are basic building blocks used to implement circuits. Across broader designs, OpenAI estimates roughly 13% lower standard-cell area for the vector unit and roughly 10% for the matrix unit. These figures concern local circuit area or latency, not whole-chip speedups.

On CPUs, or central processing units, OpenAI relays that AMD validated single-digit percentage PPA gains in critical blocks of a next-generation flagship design. OpenAI expects further GPU and CPU improvements relative to highly optimized baselines. That is an expectation for subsequent work, not an achieved result.

OpenAI also ran its own RISC-V evaluation. RISC-V is an open processor instruction set. The subjects were two open-source 32-bit cores: PicoRV32, a compact core intended for programmable logic chips, and Ibex, an embedded processor used in the OpenTitan security chip. The goal was higher instruction throughput—processing more instructions per unit of time. Designs received zero if simulation failed or their area exceeded 16 times the original core’s area. In OpenAI’s reported results, GPT-6.2 Astra had the highest average performance on both cores.

Another task supplied only written requirements and interface templates, leaving the model to choose data flow and pipeline organization. OpenAI says that at a comparable budget of about 112,000 tokens, all five models scored higher on both cores when designing from a specification than when revising the original design, with a clearer gain for GPT-6.2 Astra on Ibex. Starting from a specification, however, lets a model omit features such as performance counters and extra instructions that contribute little to the tested throughput. Some of the advantage may come from those omissions. Moving closer to production design requires retaining those features, tightening area limits and checking timing after physical synthesis, when logic is mapped to an implementation in a specific manufacturing process. These are design evaluations, not results for production CPUs.

iHolyshit review

What matters more to me is the demonstrated path connecting model-driven circuit rewrites, behavioral verification and measurement of design trade-offs. Up to roughly 14% improvement in design metrics deserves attention, but local area, latency and evaluated throughput each have their own boundaries. To me, some changes entering a design sent for manufacturing is more meaningful than treating those figures as graphics-card speedups.

Share this article

What do you think of this story?

Anonymous feedback, up to 5 reactions per article.

Reader comments

Comment without signing in, or sign in to display your member name.

Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.

0/500

No comments yet. Be the first to share.

Sources