AI DevelopmentIndustry Guide11 min readPublished

How OpenAI Helped AMD Redesign Parts of Its Flagship GPU

AMD carried model-written changes into a real manufacturing design. The reported gain of up to 14% belongs to the circuits under review. A buyer needs a different measurement to put a number on the finished processor.

DA
Digital Applied
PublishedNovember 21, 2026
Read time11 min
SourcesOpenAI · MLCommons

AMD sent a chip design for manufacturing that included circuit changes suggested by an OpenAI model. OpenAI reported the work on November 18, 2026, and put the best improvement on AMD’s flagship GPU design blocks at 14% against designs that engineers had already optimized. For a hardware team, the significant decision is the one AMD took after the model finished writing. It chose to build some of the changes.

A buyer reading the same news has a different question. Would a server using that processor handle more requests, use less electricity or cost less per completed job? The announcement gives no measured change for an entire AMD GPU. It does describe the checks the model’s work passed and, on a separate chip called Jalapeño, what engineers learned when the hardware came back. OpenAI’s November 18 chip-design report

Key takeaways
  1. 01
    AMD changed a manufacturing designSome model-proposed revisions reached the design that AMD sent to be made. Others entered production design files for later versions. OpenAI did not identify the individual circuits in either group.
  2. 02
    Keep the 14% attached to its baselineThe upper figure compares flagship GPU circuit designs with already optimized versions under power, performance and area evaluation. The report publishes no corresponding whole-GPU energy or speed measurement.
  3. 03
    Jalapeño supplies the physical exampleOpenAI estimated smaller vector and matrix units on its own inference chip and reported that the changes worked during silicon bring-up. Its detailed arithmetic examples belong to that separate processor.
  4. 04
    Buyers can specify the missing testChoose a model, workload, latency target and power measurement point. Compare complete systems on accepted work delivered under those same conditions before assigning a financial value to a circuit change.

01 — Circuit descriptionWhich part of the GPU did the model change?

Chip designers use a hardware description language to specify what happens to a value between clock edges. At the register-transfer level, often shortened to RTL, they describe where values are held, which calculations happen and when the results move on. SystemVerilog is the language OpenAI names for the AMD work. A compiler for hardware can turn that description into a collection of connected logic cells. Changing the description can change how much hardware is needed to do the same job.

Imagine a block that computes two kinds of sums. If the design guarantees the two operations never occur at the same time, one adder might serve both. That can save space. It also requires extra logic to choose its inputs. If the input selection slows the most timing-sensitive path, the shared adder may miss the clock period the design must meet. That is an illustrative design choice. AMD has not disclosed that it made this particular change.

The model had something an assistant writing a slide deck rarely gets. Engineers could evaluate each candidate using tools that understand the design and its timing constraints. A fluent explanation of why the change should work counted for little if its implementation broke the block. A proposal could also pass a functional check and still lose once the tools estimated its physical cost.

02 — First decisionDoes the circuit still deliver the right result?

Two circuits can return the same value on a routine example and disagree on a rare input or on the cycle when a downstream operation needs it. For a simple example, consider a queue that accepts new work while an older result waits for its consumer. A rewrite could preserve every arithmetic answer but lose an item when the consumer pauses. Running a few calculations separately would miss the failure. The check has to reflect the contract that neighboring blocks rely on.

OpenAI says Synopsys design tools checked that each rewrite preserved the behavior of its block. The report does not publish the underlying AMD designs, the verification settings or the list of exceptions. Readers can credit the reported verification step while leaving its precise coverage to the engineers who have the design. An integration team considering the same method can ask which assumptions are built into the check and who owns them.

One practical review question is what happens after reset, during a stall and on the first transaction after the stall clears. Those inputs are easy to omit from a demonstration yet can matter every time a chip starts or a queue fills. A reviewer who finds that a candidate depends on a stronger assumption than the original design has a concrete reason to send it back.

A question for a design review

What can a neighboring block observe, including when an output is valid? A proposed rewrite has to respect that interface before its estimated savings can influence a manufacturing decision.

03 — Second decisionWhich trade-off is worth the silicon?

Power, performance and area are usually abbreviated to PPA. Power matters for electricity use and heat. Performance might refer to the timing a circuit can meet or the work it can produce per cycle. Area matters because physical resources are finite. Improving one can change the others. Adding a register can shorten a long timing path, for example, while increasing storage and delaying the availability of an individual result by a cycle.

Latency and throughput are easy to confuse at this point. Imagine a two-stage arithmetic circuit that accepts a new input every clock cycle. One result takes two cycles to emerge, but once the stages fill, the circuit can finish one result every cycle. A one-stage replacement could reduce the wait for a single result while leaving sustained output at one per cycle. The benefit to a program depends on whether it can issue independent work or must wait for that particular result before issuing the next operation.

An engineer evaluating a candidate can hold the clock target constant and look for a smaller implementation, or keep an area budget and try to meet a tighter timing target. Those experiments answer different questions. A useful results sheet identifies the objective, the quantities held fixed and how the baseline was produced. Otherwise, two designers can honestly report an improvement while describing different engineering outcomes.

OpenAI reports an improvement of up to 14% on already optimized flagship GPU intellectual property. The announcement gives neither a split between the three measures nor a per-block table that would let a reader reconstruct the peak. Treat it as an upper result from the design exercise. Applying 14% independently to power draw, clock speed and chip area would invent three measurements from one statement. How to read disclosed benchmark comparisons

04 — Evidence by stageWhat did AMD actually send to manufacturing?

Moving a change into the design sent to a manufacturer is a consequential engineering decision. The team accepts the implementation within the constraints of that revision. Keeping another candidate in the production files also matters. Engineers can use it in a later revision, where adjacent blocks or physical constraints may differ. Neither action tells a buyer when a particular system will be available to order.

The manufacturing milestone is especially informative for teams considering whether model-assisted design can survive an established engineering process. A toy design can discard a requirement to improve a score. A team preparing a production submission has to decide whether the candidate fits a much larger design that other engineers already depend on. The public account provides evidence of that decision for some changes. It gives no count of proposals reviewed, so a percentage of successful suggestions cannot be calculated from it.

The November account does not map a named AMD block, or the best-scoring rewrite, to either disposition. In particular, it leaves open whether the candidate behind the largest reported improvement was one of the changes in the submitted design. The third row below explains why bringing Jalapeño into the same discussion can make the AMD evidence look stronger than the account supports.

OpenAI’s report on AMD and Jalapeño, November 18, 2026. The report groups the AMD changes and gives circuit-level examples for Jalapeño.
DesignStage reachedWhat that establishes
AMD flagship GPU, some rewritesIncluded in a design sent for manufacturingAMD adopted changes for that design. OpenAI gives no workload measurement on a finished AMD processor.
AMD flagship GPU, other rewritesAdded to production files for future revisionsThe implementation is available for later use. Its inclusion in a shipped revision remains unspecified.
OpenAI Jalapeño, compute-unit optimizationsChecked during bring-up of physical siliconOpenAI reports working hardware. Its area figures are engineering estimates for the named units.

05 — Physical hardwareJalapeño gives the arithmetic a unit

OpenAI’s own inference chip offers a more detailed example. In its matrix unit, a multiplier rewrite used about 15% less cell area in synthesis and cut latency from two clock cycles to one. A separate adder used about 15% less cell area and kept its two-cycle latency. Identical area percentages here describe two different timing outcomes. The first could help an operation that depends on the product. The second could free space while preserving the existing schedule.

For the larger units, the team estimated about 13% less standard-cell area in the vector unit and about 10% less in the matrix unit from model-proposed changes. A standard cell implements a small piece of logic. Its area measures a different part of the design from the entire die, which also has memories, wiring and other structures. Adding the unit percentages would also mix denominators. Each refers to its own unit before the changes.

OpenAI says it verified the optimizations during silicon bring-up, the stage when engineers power and exercise fabricated hardware. Its August account separately describes running inference workloads on Jalapeño systems and distinguishing response speed from throughput per kilowatt. Even a system result on this chip belongs to Jalapeño, its software and its workload. Transferring it to a future AMD GPU would require fresh measurements. OpenAI’s Jalapeño system results

06 — Worked exampleHow large is the block’s share of the whole?

Suppose a team has an imaginary chip with 100 equal units of cell area, and a candidate occupies 10 of those units. If the candidate shrinks by 15% while everything else stays fixed, it occupies 8.5 units and the total becomes 98.5. That is 1.5% less cell area for the chip in this simplified example. The arithmetic requires both the block’s original share and a specific area reduction. Neither input is disclosed for the AMD result.

Speed needs different arithmetic. Consider a job that spends 10 milliseconds in one calculation and 90 milliseconds waiting for data. Halving the calculation time saves five milliseconds, bringing the total to 95. It would be reasonable for the designer of the calculation to describe a twofold speedup there. A customer waiting for the complete job experiences five milliseconds less elapsed time. If another operation had been running in parallel and finished later, even that saving might disappear.

The same distinction matters in capacity planning. A service blocked on memory traffic might gain little throughput from faster arithmetic. A team constrained by a power limit might use a more efficient block to run more work inside the same power envelope. Those are workload hypotheses an operator can test. Neither supplies an estimate for AMD’s product. How inference costs change with utilization

The arithmetic changes again when there are several proposed savings. If two changes each remove the same piece of duplicated hardware, the second can look attractive in isolation and add little after the first is applied. An engineer needs a combined design built from the same baseline to measure their joint effect. For a buyer reading a list of attractive best cases, asking which changes coexist in the submitted revision can be more informative than asking for another percentage.

07 — Research contextWhat the open-core tests let a model leave out

The same research report includes experiments on PicoRV32 and Ibex, two open-source 32-bit processor cores. These give researchers access to designs they can modify and compare under a shared scoring procedure. The measured objective combined work per cycle with estimated clock speed. A simulation failure or a design exceeding the area limit earned a zero. The score describes performance within those rules.

OpenAI observed that models often scored better when starting from a specification than when revising existing core code at a comparable token budget. One explanation it offers is that a new design can omit features such as performance counters or extra instructions that the test workloads use little. A product team may need those very features to diagnose a fault or to support customer software. An omitted feature can improve a score while increasing the cost of adopting the design.

The lesson for an engineering lead is to put the expensive obligations into the acceptance test. Require the instructions that shipped software uses, describe the counters that support engineers rely on and set a realistic area limit before comparing candidates. That makes a benchmark more representative even if the model’s score falls. This open-core experiment supplies context on evaluation design. AMD’s manufacturing decision stands on its own evidence. OpenAI’s description of the RISC-V evaluations

08 — Buying decisionWhich measurement would change the order?

Start with the decision the purchase has to support. An inference service might need to handle a fixed number of simultaneous conversations while keeping the slowest acceptable reply within a deadline. Specify the model, its precision, the sizes of the inputs and outputs and the software versions that will serve it. Make the existing system run the same work. A lower electricity figure at a much lower request rate is a poor basis for an operating budget.

Ask where power was measured and which equipment is in the total. Card power and server input power have different boundaries. Ask how many accepted outputs the system delivered during that same interval and how slow the busiest requests became. For an example of a structured comparison, MLCommons defines its inference system in terms of hardware and software and uses scenario-specific quality and latency requirements. MLCommons inference rules

For an illustrative power comparison, suppose two servers each draw an average of one kilowatt during an hour-long run. Each consumes one kilowatt-hour. If the first completes 1,000 accepted jobs and the second completes 1,250 at comparable response times, the energy per job is 1.0 and 0.8 watt-hours respectively. The second completes 25% more work and uses 20% less energy per job under those assumptions. Neither percentage can be derived from a designer’s PPA figure. The calculation shows exactly which counters an operator needs to collect.

There is a reason to record the AMD news in a supplier assessment today. A vendor has described model-generated changes surviving verification and moving into a manufacturing design. That is evidence about the engineering process. A capacity forecast can retain its existing throughput and energy assumptions until the supplier provides measurements on the configuration the buyer can actually deploy. How custom silicon fits a cloud hardware decision

For a team comparing hosted inference with its own hardware, the same workload specification can be reused across both trials. A cloud invoice supplies a price for delivered work. A hardware trial supplies the utilization and energy inputs for an ownership estimate. Digital Applied’s AI transformation services

Engineering capability
Request the baseline RTL, acceptance criteria and a before-and-after result at matched constraints, under confidentiality if needed.
Circuit review
Server capacity
Measure accepted work per second at the target latency, precision and concurrency on the proposed system.
Workload test
Operating cost
Price the equipment, energy and utilization needed for the required output. Record the power measurement boundary.
Cost per job
Procurement

Keep the result at the level where it was measured

AMD’s decision to manufacture a design containing model-written changes is worth reporting on its own terms. For a purchase order, the missing figures are work delivered, latency and energy at the operating point you expect to use. A vendor can answer those questions with a system benchmark once the relevant hardware is available.

AI infrastructure

Compare the cost of serving your actual workload

Digital Applied can help define the request mix, latency threshold and cost measures for comparing inference options.

Representative requestsComparable measurementsOperating cost
Evaluation inputs

Bring the current service baseline

  • →Model, precision and serving software
  • →Traffic volume and response-time targets
  • →Measured usage and supplier pricing
Answers

Frequently asked questions

OpenAI describes rewriting parts of the SystemVerilog circuit description for AMD’s flagship GPU. AMD selected some changes for a design sent to manufacturing and retained others for future versions. The announcement does not identify a full GPU designed from a blank specification.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

AI Development

What GPT-Live 1.1 Charges While a Voice Call Is Silent

A ten-minute GPT-Live-1.1 call can cost $0.50 or $0.38 in voice charges. See when WebRTC standby starts, how tool waits are billed and when to hang up.

November 21, 2026 · 11 minRead
AI Development

Prompting Mistakes OpenAI, Anthropic and Google Warn Against

The prompting mistakes each vendor's own guide names, quoted: thinking prompts on reasoning models, shouting, prefills, examples, and where the three disagree.

November 1, 2026 · 10 minRead
AI Development

GPT-5 and o3 Leave OpenAI's API on Dec 11: Migration Guide

OpenAI shuts down gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro on December 11, 2026. The snapshots, the named replacements and what the switch costs.

October 31, 2026 · 10 minRead
AI Development

GPT-6.1 Terra: Pricing, Computer Use and Upgrade Guide

GPT-6.1 Terra lowers API prices for visual work and computer use. Compare its real cost, long-context rates, benchmark evidence and migration requirements.

October 27, 2026 · 9 minRead
AI Development

AI Agent Governance: Policy and Compliance 2026 Guide

AI agent governance framework for enterprises — access control, audit trails, data residency, and compliance with EU AI Act and SOC 2 requirements.

May 23, 2026 · 20 minRead
AI Development

Google AI Plans: Free vs Plus vs Pro vs Ultra 2026

Google's AI subscription tiers after I/O 2026 — AI Plus $7.99, AI Pro $19.99, AI Ultra $100 (new), AI Ultra $200 (was $250). Feature matrix and decision tree.

May 23, 2026 · 14 minRead
Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source