The Next Era of Semiconductors

For the past 60 years, semiconductor progress has been powered by several mutually reinforcing forces, from innovations in integrated circuit technology and architecture all the way to advances in chip design tools, verification, manufacturing test and field reliability.

Most importantly, these advances compounded, helping build the modern computing world: the PC, the internet, the smartphone, the cloud, and now AI.

Today, many of those curves are slowing at the same time, and some are beginning to bend in the wrong direction.

The old transistor scaling magic is fading. Well-known architectural tricks alone are no longer adequate. Building enormously complex systems requires extraordinary effort.

As those gains diminish, extracting new ones entails inexorably greater complexity. But how do we know the resulting systems actually work?

Verification is already struggling. Manufacturing test is under growing pressure. Ensuring reliable operation is getting harder. Our ability to create complex designs is increasingly outpacing our ability to establish their correctness.

This isn’t all gloom and doom. There are extraordinary innovations ahead: new memory and logic technologies, dense and heterogeneous 3D integration, integrated photonics, co-optimized algorithms and architectures, in-field scan testing, and much more.

Importantly, AI has arrived with almost perfect timing.

The obvious story is seductive:

AI will design the chips. AI will write the RTL. AI will derive the best architecture. AI will verify everything. AI will debug failures. AI will make semiconductor engineers 10x, 100x, maybe even 1,000x more productive.

But that story may also be naive:

If AI creates accelerator designs 100x faster but we can’t verify them 100x faster, then we have simply amplified the verification bottleneck.

If AI can discover 3D architectures no human would have imagined, but we can’t simply trust them, then “more intelligence” doesn’t automatically imply more progress.

If current chip design abstractions are reaching their limits, then teaching AI to use them dramatically faster would be like putting a jet engine on a horse carriage.

For the first time in decades, some of the fundamental questions in chip design and verification are back on the table. For example: How does “intent” become silicon? What should humans actually specify? How do we establish correctness when machines can generate designs no human could fully inspect? What becomes possible if we finally break the verification bottleneck?

That means there is something huge to invent. AI will certainly be part of the answer, perhaps a huge part. But the opportunity is much bigger than “AI for semiconductors” or “AI writes RTL.”

For 60 years, semiconductor engineers built an extraordinarily successful playbook and kept extending it. Now it is under pressure. The next one has yet to be written. That is what makes this a new beginning.

We’re excited to build what comes next. Stay tuned!