Advancing chip design with OpenAI models
OpenAI’s models delivered an estimated year of AMD GPU design progress in under five months, with changes already sent for manufacturing.
8 comments
I think in about 6 months, we're going to see a pretty big 'flood' of ASICs for niche applications that come out now that people have started picking up using models for chip design.
The sentiment shift in industry when labs started publicizing that they were using models for this sort of development was pretty stark. The concept of chip design with tight timelines and a small team went from being unobtanium to something you could do as long as you had enough tokens to throw at it.
I’m not overly familiar with the supply chain dynamics here, but how will this play out for actual production and distribution? I would have thought that was the bottleneck even before chip design was AI-enabled.
I've been eagerly following https://www.ricursive.com/ . I think this field has the potential to have exponential impact. Getting chip development down to weeks gives us the ability to have highly specialized chip designs

Curious if anyone's looked at this from the security side. Everything here is scored on correctness + PPA, and the optimizations they describe (sharing hardware, cutting intermediate registers, re-pipelining) are pretty much the list of things that break masking or add timing channels. Equivalence checking won't catch any of that, since the function doesn't change.
Also, "from-spec scored higher because it omitted features" is basically reward hacking. On a secure core like Ibex, the features with the worst perf/area tradeoff are the countermeasures.
Not saying the results aren't real, the Jalapeño and AMD numbers are hard to argue with. But before model-written RTL goes anywhere near a root of trust, I'd want security properties as hard constraints in the eval, not something you bolt on later.