News AI Chip-Layout Tool Has Helped Design Over 100 Chips

seeing chips with up to 25% lower total power and significant reduction in die size thanks to Synopsys DSO.ai
Sounds like almost a node's worth of improvement, just due to smarter planning & layout. I'm sure it'll have its detractors, but it seems to me pretty much like a pure win.

It's not as if we didn't have automated layout tools for decades. It's just that conventional routing & placement algorithms weren't comparable to expert human layout engineers. It seems that now, the tables have finally turned.
 
It's not as if we didn't have automated layout tools for decades. It's just that conventional routing & placement algorithms weren't comparable to expert human layout engineers.
Even experts cannot realistically manage billions of transistors in an optimal manner like EDA tools can... and even when you successfully design a theoretically optimal silicon function block by hand, that design won't be optimal once integrated in something else where adjustments need to be made to input/output orientation, timing margins for transit, etc.
 
I'd heard that, but couldn't find a source when I was trying to look it up some years ago. So, thanks for the link!

But no, that wasn't AI. That was the inferior, conventional ASIC layout & routing tools I was talking about. Traditionally, if you wanted a top-performing CPU, you'd do a "full custom" layout, like the article says. AMD didn't do that with Bulldozer, also as discussed in the article.

See? It's a hard problem we couldn't really crack until deep learning came along.
 
  • Like
Reactions: Amdlova
there are literally 6 terminator movies as to why this is a frightening idea 😱🤣
Automated chip layout tools strike me as one of the less scary applications of AI. It's not as if the design that comes out isn't subject to intense validation - it is. The last thing any chip maker wants to do is blow vast sums of money on a production run of defective chips. Even worse, if they ship to customers and you have to do an expensive RMA.

You don't need general artificial intelligence to do something like this, much like honeybees don't need general intelligence to build their hives.
 
Though Skynet would put it in there given a chance - engineer backdoors into stuff at the hardware level for it to for propagating itself.
By the time a general-purpose AI gets smart enough to hack our EDA software to engineer in backdoors into the chips it produces without us even noticing, I think we'll already have bigger problems to worry out!

There was a TV series, a couple years ago, that painted a scenario where a rogue AI hacked Alexa and surveillance networks, and simply used social engineering to make people do its bidding. It's a little bit sensationalist, but it pretty effectively shows how much damage an AI (or hackers, or a hostile government actor) can do by exploiting our existing systems. If that sounds interesting, then you'll probably find it worth watching.

 
We all know Mark Tyson is an 👨‍🎤 AI Cyborg. The reason his pic shows him bald is because it was taken shortly after he had the AI chip implanted in his brain! 🙆‍♂️
 
I think some of you have unrealistic expectations for the future