Article

AI scored better designing small processors than improving them

For hardware teams, OpenAI’s processor study shows why choosing a new design or modifying an existing core starts with deciding what it must keep.

Editorial illustration for AI scored better designing small processors than improving them: a document represents the research briefing. Not documentary evidence.

OpenAI reported on November 18, 2026, that all five tested AI models scored higher designing both small processor cores from requirements than modifying existing designs, at roughly 112,000 tokens per run.

For hardware teams, the useful distinction is whether the job calls for a new blueprint or a compatible replacement. The practical recommendation is to explore a fresh design when its features can be deliberately chosen, and constrain changes when existing software and connections must keep working.

That is an engineering judgment, not a rule that starting over always wins. Renovating a house and drawing a new floor plan are both design work. They are not the same commission.

What the models were asked to build

A core is the part of a processor that executes instructions. PicoRV32 and Ibex are small, 32-bit RISC-V cores for embedded control tasks, not mainstream laptop CPUs.

OpenAI supplied interface templates alongside the requirements. Both tasks shared correctness checks, memory/cache settings and throughput objectives. The comparison therefore had meaningful controls; it was not invention on an empty page.

Scores combined simulated correctness and estimated throughput. Failed runs and designs exceeding 16 times the original circuit area scored zero. That ceiling was not every design’s actual size. Averages included failures, so higher scores can reflect reliability as well as speed—not measured silicon performance.

OpenAI suggests omitted counters or extra instructions might partly explain the advantage, without isolating the cause. Physical timing and complete feature equivalence remain unestablished. These are OpenAI’s findings, without established independent replication.

A small processor may be small on purpose

The original design’s priorities matter. PicoRV32’s documentation emphasizes compact circuitry and high maximum clock frequency rather than overall performance.

Clock frequency is how quickly the hardware ticks; throughput is how much instruction work it completes. Faster ticking and more completed work are not interchangeable. A design brief must say which outcome matters, and what space is available to achieve it.

Imagine commissioning a compact controller and receiving a bulkier design that finishes a test sooner. It might be an interesting alternative, but a poor replacement if fitting into limited space was the reason for choosing the original. That is an illustrative product tradeoff, not an observed outcome of OpenAI’s designs.

The same reasoning works in the other direction. If the product can accommodate a different layout and has no need for some existing functions, preserving every old decision could unnecessarily narrow the search. Architectural exploration is useful precisely because a team can reconsider those choices deliberately.

A counter can be optional until your software needs it

Consider a less glamorous feature: a performance counter. PicoRV32’s README says disabling its ENABLE_COUNTERS option makes the documented counter-reading instructions trigger a hardware trap—the response to an unsupported instruction.

Suppose a speed test never reads that counter, but a device’s diagnostic software does. A design could satisfy the test’s needs without satisfying the diagnostic program’s. This is a hypothetical compatibility example grounded in the documentation, not a reported failure of an AI-generated core.

Ibex’s counter documentation makes their purpose clearer. Its cycle and completed-instruction counters are always available. Additional event counters, absent by default, can track matters such as time spent waiting for memory or arithmetic operations.

Those instruments need not make the current workload faster to be useful. Knowing where execution stalls can help an engineer decide what to change next. Removing them is therefore a choice about the ability to diagnose the design, not simply a way of clearing clutter.

The same speed score can describe different products

Ibex supplies a concrete illustration outside OpenAI’s experiment. In the maintainers’ configuration table, maxperf and maxperf-pmp-bmfull both score 3.13 CoreMark/MHz, a benchmark result adjusted for clock frequency. Their Yosys circuit-area estimates are 32.48 and 66.02 kGE, respectively; kGE means thousands of equivalent logic gates.

The illustrative calculation is 66.02 ÷ 32.48 ≈ 2.03: roughly twice the estimated circuit area for the same listed speed score. Both figures use the README’s Yosys column. They are not measured chip sizes or manufacturing prices, and the comparison does not isolate the cost of any single feature.

The larger configuration also offers additional instruction support and memory-protection features. A purchasing or engineering decision cannot treat those capabilities as invisible merely because one benchmark gives the same number. The right question is whether the intended product needs them.

These upstream documents illustrate real choices; they do not identify the study’s exact configurations or what each generated design omitted.

Choose what must survive before choosing the approach

For an exploratory brief, explicitly identify which functions may disappear and which interfaces must stay. For a replacement brief, start with the instructions, interrupts, diagnostic functions and connections used by the existing system. Then specify physical limits such as area, timing and power.

This is a proposed way to commission the work, not a decision rule validated by the experiment. Editing an existing core does not itself guarantee compatibility, just as a fresh design does not establish incompatibility. Both need to demonstrate the behavior the product requires.

The core experiments are separate from OpenAI’s reported Jalapeño and AMD production-block improvements. For the distinct business question behind custom AI hardware, see Jalapeño and the economics of inference.

A candidate design also faces the separate step of engineering and manufacturing review. But the immediate decision comes earlier: what may change, and what must still work? Write that down before asking AI to improve the blueprint—or draw another one.