The development
What happened
OpenAI has published a technical update on using AI models to rewrite hardware-description code and explore processor designs. The 18 November report describes circuit improvements in its Jalapeño inference chip and AMD designs, alongside separate simulation-based RISC-V evaluations. The distinction matters: OpenAI reports some production-design gains, but its processor-test scores are estimates under permissive constraints, not measured performance of newly manufactured chips. This extends earlier Jalapeño work rather than announcing a new chip shipment.
Key facts
OpenAI estimates that model-proposed changes cut standard-cell area by about 13% in Jalapeño's vector unit and 10% in its matrix unit, and says these optimizations were proven during silicon bring-up. These are block-level figures, not whole-chip savings.
For AMD GPU intellectual-property blocks, OpenAI reports up to 14% improvement in power, performance and area against heavily optimized baselines. Some changes entered a design sent for manufacturing; others are in files for future versions, not necessarily shipped products.
The models rewrite SystemVerilog RTL, which describes circuits. OpenAI says Synopsys tools checked preserved behavior and assessed trade-offs; it also reports AMD-verified single-digit percentage gains in key blocks of a forthcoming CPU, not an independently checked whole-CPU speedup.
Separate tests used two open-source 32-bit RISC-V cores, PicoRV32 and Ibex. Simulation failures or designs exceeding 16 times the original area scored zero. Scores combined hidden-workload instructions per clock with Yosys-estimated clock frequency and were normalized per core.
At roughly 112,000 tokens per run, GPT-6.2 Astra led both cores on average. All five tested models scored higher when designing from written requirements than modifying existing RTL. The 9,881 scored runs span both tasks, both cores and all tested token budgets.
OpenAI notes that designs from specifications could omit counters or extra instructions. Requiring those features, tightening area limits and checking timing after physical synthesis would make the tests more representative. Its GPT-Synopsys training partnership is not a public release.
The context
Why it matters
TideReach assessment: AI-assisted circuit optimization and open-ended processor design are different claims. The former can be checked against an existing block's behavior and physical-design trade-offs; the latter depends heavily on the specification, validation and area budget. OpenAI's account suggests useful engineering assistance, but the simulated scores do not establish autonomous production-ready chip design, a universal hardware speedup or independent replication of the reported commercial gains.