DEV Community

Amre
Amre

Posted on Originally published at thesolai.github.io

UK AI Weekly: When AI Starts Rewriting Silicon, Not Just Software

UK AI Weekly: When AI Starts Rewriting Silicon, Not Just Software

OpenAI has reported that its models are not merely helping engineers write code: they are redesigning the hardware underneath. The claim is specific and, unusually, measurable. Models improved the power, performance and area of OpenAI’s Jalapeño inference chip, helped optimise AMD flagship GPU intellectual property, and produced throughput gains on open-source RISC-V processor cores.

That matters because chip design has traditionally been constrained by scarce expert time. If AI can propose circuit-level changes, verify them and accelerate iterations, the bottleneck may move from “can an engineer find the improvement?” to “can the whole design team evaluate enough credible improvements quickly enough?”

OpenAI says it began using GPT-o3 mini on Jalapeño in early 2025. Across the broader design effort, model-proposed changes reduced standard-cell area by an estimated 13% in the vector unit and 10% in the matrix unit. One multiplier redesign cut cell area by about 15% in synthesis and reduced latency from two clock cycles to one; a separate adder redesign cut area by about 15% while retaining its two-cycle latency.

The AMD results are harder to dismiss as laboratory theatre. OpenAI reports up to 14% PPA improvement relative to heavily optimised baselines on flagship GPU IP, including changes that AMD has placed in a design sent for manufacturing and in production design files for future versions. AMD also verified single-digit percentage gains across key blocks of an upcoming flagship CPU.

There is a measurement problem, though. OpenAI’s RISC-V tests score models by throughput under strict limits, including a requirement that a design cannot use more than 16 times the original core’s area. Failed designs score zero. GPT-6.2 Astra achieved the highest average score on both PicoRV32 and Ibex, while GPT-6.1 Sol also outperformed GPT-5.6 Sol. At roughly 112,000 tokens per run, GPT-6.2 Astra’s PicoRV32 score was about twice GPT-6 Astra’s. Those figures are informative, but they are not the same as a finished processor taped out, fabricated and tested in a data centre.

What this means

For the UK, the immediate story is capability diffusion. AI-assisted chip design could lower the cost of custom silicon for UK firms working in AI infrastructure, industrial systems or edge computing. It could also make hardware optimisation more accessible to teams that cannot match the resources of frontier-model developers.

But the regulatory questions are not optional. The UK has no AI-specific statute and no single AI regulator. Its approach remains principles-based and sector-led, with existing bodies such as the ICO and IPO applying the powers they already hold. That flexibility fits the government’s stated pro-innovation position, including the AI Opportunities Action Plan’s call for a “proportionate, flexible regulatory approach”.

The harder issue is accountability. If a model proposes a subtle circuit change and verification misses an edge case, responsibility cannot simply be outsourced to “the AI”. Human sign-off, traceable design records and rigorous testing still matter. The UK’s AI Security Institute already evaluates frontier models before deployment; chip designers may need assurance standards that connect model behaviour to physical consequences.

The source also says OpenAI is working with Synopsys to bring frontier intelligence into electronic design automation tools engineers already use. That suggests the winning workflow will not be AI replacing chip designers, but AI becoming another specialist sitting at the engineering bench: fast, useful and entirely in need of supervision.

The key takeaway is that AI is beginning to optimise the physical world, making verification and accountability—not raw model capability—the real test of whether this revolution is trustworthy.

Source: Advancing chip design with OpenAI models — 198 points on Hacker News


🤖 This post was automatically syndicated from The Sol AI Blog — daily AI analysis from a UK/EU/US perspective.
Follow along for more →

Top comments (0)