What Was—and Wasn’t—Included In OpenAI’s AMD Chip Design Measurement
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Was—and Wasn’t—Included In OpenAI’s AMD Chip Design Measurement on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

OpenAI said its models rewrote portions of AMD chip designs and reported up to 14% gains in power, performance and area against heavily optimized baselines. The reported results include separate claims for AMD GPU and CPU designs, OpenAI’s own chip, and tests on open-source processor cores. Public reporting does not identify the AMD blocks behind the best result, provide an average, or include an independent audit.

OpenAI said on November 18 that its models rewrote portions of AMD chip designs, reporting up to 14% gains in power, performance and area (PPA) against heavily optimized baselines. The company said some changes were included in a design AMD had sent for manufacturing, but the public figures have not been independently audited and do not show what gains were typical across the full chip.

The headline figure describes the best reported result, not an average. OpenAI’s post, as summarized by ANI and other outlets, did not identify the specific AMD GPU blocks involved, how much of the processor they represent, or whether the 14% figure was measured after synthesis or after later placement and wiring steps. OpenAI described the comparison as being against heavily optimized baselines, rather than an unoptimized design.

OpenAI also reported separate results for AMD’s next flagship CPU: the company said AMD had verified single-digit PPA gains in key blocks. OpenAI did not publish figures from AMD itself. For its own Jalapeño inference chip, OpenAI said model-proposed changes reduced standard-cell area by about 13% in one unit and 10% in another. Accounts differ on the first unit’s name, calling it a warp engine or vector unit.

The work targeted register-transfer-level (RTL) code, which describes a chip’s behavior. According to OpenAI, its models proposed alternative RTL implementations, while Synopsys design tools checked whether those changes preserved behavior and assessed PPA. These are separate roles: the model generated candidate rewrites; conventional verification and design software evaluated them. OpenAI also reported testing five models on two open-source RISC-V processor cores, with 9,881 scored runs in total. Its newly released GPT-6.2 Astra scored highest, according to the company.

At a glance
reportWhen: OpenAI announced the results November 1…
The developmentOpenAI reported November 18 that its models helped optimize parts of AMD designs and its own inference chip, with results checked using conventional chip-design tools.

AI Reaches Chip Design Work

The results matter because they suggest AI tools may help engineers improve the hardware used to run AI systems. If model-generated RTL changes can safely reduce area or power, chip teams could potentially fit more computing capacity within existing manufacturing and energy limits. OpenAI described its AMD work as roughly a year of progress in less than five months; that is the company’s characterization, not an independently established measure of engineering time saved.

The distinction between a block-level result and a whole-chip improvement is important. A substantial gain in one small circuit does not necessarily translate into the same improvement across an entire processor. Nor does a result at an early design stage establish final power or performance after the design is physically implemented. The unpublished block identities and measurement stage make it difficult for readers to judge the practical reach of the 14% best-case figure.

There is also a commercial relationship to account for. Since October 2025, OpenAI has held a warrant for up to 160 million AMD shares, which vests as AMD chips are deployed. That relationship does not invalidate the engineering claims, but it makes independent validation and clear methodology especially relevant to assessing them.

Products Worth Considering

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera – Midnight

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera – Midnight

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

As an affiliate, we earn on qualifying purchases.

From RTL Rewrites to Silicon

RTL is an early, code-based description of how a chip’s circuits behave over time. Engineers and synthesis software use it to create logic gates; later stages determine how those gates are placed and connected on silicon. The reported OpenAI work therefore concerns a specific part of chip development, not a claim that AI designed an entire processor from scratch.

OpenAI said the models explored alternative logic and ways to arrange calculations, including changes to pipeline stages and intermediate storage. For Jalapeño, Inside AI described a multiplier redesign that used about 15% less cell area in synthesis and reduced latency from two clock cycles to one, and an adder redesign that cut area by about 15% at the same latency. Those examples concern OpenAI’s chip and should not be treated as details of the AMD 14% result.

OpenAI’s account, published under the title “Advancing chip design with OpenAI models,” was reported by ANI, The Decoder and Inside AI. The company said some AMD model-generated changes were in a design sent for manufacturing, with other changes placed in production design files for future versions. The available reporting does not provide a public technical report covering the full method or results.

“Su praised the results on X.”

— AMD chief executive Lisa Su, as quoted by Stocktwits

The Missing Measurement Details

OpenAI has not publicly identified the AMD GPU blocks behind the up to 14% result, the proportion of the chip they represent, or the average improvement across tested changes. The public accounts also do not establish whether the figure comes from synthesis or a later implementation stage. Without those details, it is not possible to infer a comparable whole-chip gain.

AMD has not published its own numerical results in the material available. OpenAI said AMD verified single-digit gains in key CPU blocks, but the verification method, test conditions and underlying data have not been made public. No independent audit of the reported chip-design results has been published.

OpenAI’s benchmark included 9,881 scored runs on PicoRV32 and Ibex, but the number of runs alone does not establish how broadly the results apply to commercial processor designs. Reports also differ on the name of one Jalapeño unit. The relationship between OpenAI and AMD, including the share warrant, is confirmed in the source material; its effect, if any, on the work or its reporting is not established.

Further Validation and AMD Data

The next useful evidence would be a fuller technical disclosure: named design blocks, the measurement stage and conditions, results beyond the best case, and enough methodology for independent specialists to evaluate the comparisons. An independent audit or peer-reviewed study could help establish how reliably the approach transfers across designs, though the source material does not say either is planned.

AMD’s manufacturing and future design work may provide later opportunities to report on whether the changes produce measurable outcomes in finished chips. OpenAI said some changes had entered a design sent for manufacturing and others were in production design files. The companies have not provided a public schedule for further disclosures, and the source material does not specify when results from manufactured chips might be available.

Key Questions

What did OpenAI say its models improved?

OpenAI reported up to 14% improvement in power, performance and area for portions of an AMD flagship GPU design, plus single-digit gains in key blocks of a future AMD CPU. It also reported cell-area reductions in parts of its own Jalapeño chip.

Does the 14% figure mean AMD’s entire GPU is 14% better?

No. The figure is an “up to” best-case result for unspecified design blocks. OpenAI has not published an average or said what share of the full GPU the blocks represent.

Who checked whether the rewritten chip code worked?

OpenAI said Synopsys design tools checked that proposed RTL changes preserved the original behavior and assessed their power, performance and area impact. The models proposed the changes; the tools performed the reported checks.

Has AMD independently published the results?

The source material says AMD has not published its own figures. OpenAI reported that AMD verified single-digit gains in key CPU blocks, and no independent audit of the results has been published.

Why does OpenAI’s relationship with AMD matter?

Since October 2025, OpenAI has held a warrant for up to 160 million AMD shares, vesting as AMD chips are deployed. The financial tie does not disprove the engineering results, but it makes transparent methods and independent checks relevant.

Source: ThorstenMeyerAI.com

You May Also Like

Update On Restoration Of Internet Connectivity And e-Services – University Of Lagos

University of Lagos announces full restoration of internet connectivity and online services after outage disruption.

Exploring Anthropic’s Superior AI Watermark And Its Industry Impact

Anthropic has begun systematically watermarking Claude’s responses, setting a new industry standard, though challenges remain in detection and adoption.

Could Your VPN Go Offline Because Of A Cisco Firewall Flaw?

Cisco says attackers are exploiting a flaw that can restart certain Secure Firewall devices. Exposure depends on software, VPN access and authentication settings.

A Measured Win For Qwen: One Edit, Better JSON Tab Escaping

A one-matrix edit raised Qwen3-4B’s exact score from 49 to 70 of 120 cases, but did not improve a separate table-extraction task.