NVIDIA PrefixRL Model Designs 25% Smaller Circuits, Making GPUs More Efficient

[Valantar]

Tbh, I don't think I've seen any other major chipmaker confirm that they're actually using AI-generated layouts in high volume, large size chips before. I might obviously be wrong, but all I can remember seeing is various reports of this being tested. That also clearly leaves the possibility of it being used without this being public.

 

[Wirko]

They didn't explain exactly what was reduced. Probably a combination of less wasted space between transistors, fewer transistors overall, and fewer large transistors (those that need to be larger because they need to drive larger currents).

 

[Punkenjoy]

Better than what? It's better than the classic attempts by a measurable 25%. Better than the perfect design? Probably not.

You said Nvidia use a better model from the time being, i just wanted to know why you were saying that. So Better model than a classic attempt? or Better model than other competitors ai model ?

 

[bug]

You said Nvidia use a better model from the time being, i just wanted to know why you were saying that. So Better model than a classic attempt? or Better model than other competitors ai model ?

Well, whatever they use, has to be better, since it yields something smaller than everybody else.
Like @Valantar said, others have not confirmed using AI, so I'm sure there are neural networks or something like that to compare. But the solution overall is clearly better.

 

[R0H1T]

It just happens Nvidia employs a better model, for the time being.

You sure about that? IIRC Google has a similar DL solution & they've shown great results as well. It was used by AMD as well I think?

Google claims it is using A.I. to design chips faster than humans

Google claims that it has developed artificial intelligence software that can design computer chip "floorplans" faster than humans can.

www.cnbc.com

 

[Valantar]

You sure about that? IIRC Google has a similar DL solution & they've shown great results as well. It was used by AMD as well I think?

Google claims it is using A.I. to design chips faster than humans

Google claims that it has developed artificial intelligence software that can design computer chip "floorplans" faster than humans can.

www.cnbc.com

Yes, that has been reported, but only in experimental applications. Has it actually been reported to have seen use in real world chip designs?

 

[R0H1T]

Of course, it's right there ~

“Our method has been used in production to design the next generation of Google TPU,” wrote the authors of the paper, led by Google’s co-heads of machine learning for systems, Azalia Mirhoseini and Anna Goldie.

 

[Punkenjoy]

Well, whatever they use, has to be better, since it yields something smaller than everybody else.
Like @Valantar said, others have not confirmed using AI, so I'm sure there are neural networks or something like that to compare. But the solution overall is clearly better.

Well it yields something better than most common vendor, we don't know exactly what it look like versus what Intel, AMD, Qualcomm, Apple etc. use right now. They probably all their own in house way of doing it.

 

[Valantar]

Of course, it's right there ~

Interesting! Still, that's a much lower volume and smaller scale chip than a large datacenter GPU.

 

[Wirko]

Hard to say. At what point does reinforcement learning become "deep"? It's possible others also use reinforcement learning (at least in some aspect), but treat it like a trade secret and don't talk about it. It's also possible Nvidia is the first that made this work.
What I meant to say is that I am pretty certain, in some form or another, others also use some AI techniques in their product pipelines.

Details sure aren't public but everyone is using reinforcement learning, it's not something they would want to hide.

Synopsys, Cadence, Google And NVIDIA All Agree: Use AI To Help Design Chips

Synopsys started the buzz in 2020, and now Google, NVIDIA, and Cadence Design have joined the party.

www.forbes.com

 

[bug]

Of course, it's right there ~

Like I said, it's pretty much impossible for computers to figure out simpler things than designing a CPU (http://www.iaeng.org/publication/IMECS2008/IMECS2008_pp1897-1901.pdf).
Some sort of "AI" has been employed by these problems for decades.

What we should be concentrating on, is what exactly passes as "AI". Years ago, AI was defined as something able to pass the Turing test (https://en.wikipedia.org/wiki/Turing_test). It needed to learn and adapt. In the past few years, "AI" means a neural network, deeper than 3 or 4 layers (hence deep-everything). It learns nothing once deployed, but it can be trained further and deployed again. And while that may yet qualify as a passable definition of AI (by very, very lax standards), once AI has become a buzzword, now almost every program that an if in it, claims to be employing "AI".
So you see, it's if other have been employing AI, imho, it's determining what they mean by AI when they said they used it.

 

[Wirko]

What we should be concentrating on, is what exactly passes as "AI". Years ago, AI was defined as something able to pass the Turing test (https://en.wikipedia.org/wiki/Turing_test). It needed to learn and adapt. In the past few years, "AI" means a neural network, deeper than 3 or 4 layers (hence deep-everything). It learns nothing once deployed, but it can be trained further and deployed again. And while that may yet qualify as a passable definition of AI (by very, very lax standards), once AI has become a buzzword, now almost every program that an if in it, claims to be employing "AI".
So you see, it's if other have been employing AI, imho, it's determining what they mean by AI when they said they used it.

In lack of a good definition, this one is at least easy to memorise:

https://twitter.com/i/web/status/1065778379612282885

Difference between machine learning and AI:
If it is written in Python, it's probably machine learning
If it is written in PowerPoint, it's probably AI

 

[Panther_Seraphin]

And don't forget the most important aspect (for Ngreedia).....

HIGHER PRICES !​

Yes I know R&D aint cheap, but this all sounds like just ANUTHA way to justify keeping GPU prices & profits at scalper/pandemic levels, which they have become addicted to like crackheads & their rocks... they always need/want moar and can't quit even if they wanted to....

More like keep current high prices while being able to be reduce costs/decrease die size increasing yields and increasing more dies per wafer.

 

[Steevo]

I would wager they trained it to look out for standing wave resonance in traces, induction lessening the need for capacitors, harmonic interference from nearby traces or switching speed, gate voltage roll off and all of the fun stuff that usually gets traces moved away from each other to prevent issues at design speed/TDP.

there are formulas for all of this but the actual chip may have required a few spins to get on target and this AI may be able to do it in software and also seems to eliminate the wasted space due to design flaws.

 

[R-T-B]

That's mainly because they're centre aligned. If they were corner aligned instead, the difference would be a lot more intuitive, even if it still doesn't "look like 25% less".
View attachment 254382

That's also a single adder circuit, not a complete die.

I know AMD does it, and like you, pretty sure Intel and other manufacturer use AI too. But what make you think Nvidia use a better model ?

The article we just read?

So Better model than a classic attempt?

There is no "classic attempt." All manufacturers use AI-driven algos to design circuits. This is in the article above.

Details sure aren't public but everyone is using reinforcement learning, it's not something they would want to hide.

Synopsys, Cadence, Google And NVIDIA All Agree: Use AI To Help Design Chips

Synopsys started the buzz in 2020, and now Google, NVIDIA, and Cadence Design have joined the party.

www.forbes.com

I mean the article literally says this too. My god the arguments we'd save if people read the whole article.

 

[Punkenjoy]

That's also a single adder circuit, not a complete die.


The article we just read?


There is no "classic attempt." All manufacturers use AI-driven algos to design circuits. This is in the article above.


I mean the article literally says this too. My god the arguments we'd save if people read the whole article.

So the article say Intel, Apple, AMD and others big chip manufacturer use the 2 tools listed in the article ? I don't think so. They could use those, but it's not something declared in the article

 

[R-T-B]

So the article say Intel, Apple, AMD and others big chip manufacturer use the 2 tools listed in the article ? I don't think so. They could use those, but it's not something declared in the article

Do you need me to literally quote it for you?

NVIDIA uses the PrefixRL model to outperform traditional EDA (Electronics Design Automation) tools from major vendors such as Cadence, Synopsys, or Siemens/Mentor. EDA vendors usually implement their in-house AI solution to silicon placement and routing (PnR); however, NVIDIA's PrefixRL solution seems to be doing wonders in the company's workflow.

Of course some huge companies may use something different, but this is the industry norm and it's all AI driven.

 

[Punkenjoy]

Do you need me to literally quote it for you?


Of course some huge companies may use something different, but this is the industry norm and it's all AI driven.

Yes, i never said they do not use AI. From what i know, they all declared they do. What i say is about this line:

NVIDIA uses the PrefixRL model to outperform traditional EDA (Electronics Design Automation) tools from major vendors such as Cadence, Synopsys, or Siemens/Mentor. EDA vendors usually implement their in-house AI solution to silicon placement and routing (PnR); however, NVIDIA's PrefixRL solution seems to be doing wonders in the company's workflow.

My point is Nvidia is saying beating these vendors but we do not know how it stack against what other chip maker use. They can/could use their own ai model internally like Nvidia do. So it's nice for Nvidia but it's an incomplete picture and people should remain cautious of the signification of this claim.

 

[R0H1T]

Well here's the thing, since they're all competing with each other they won't publish like for like results against each other's (AI driven) methods not unless they have a spy & they announce this to the world showing the comparison