Google AI designs* chip in 6 hours

[WolframaticAlpha]

Summary

Google has designed an artificial intelligence program that can design a chip 'floorplan' in 6 hours, something that takes months if it is done by human engineers. The floorplan of a chipset is the basic design of a silicon system-on-chip that determines the optimal positioning of the CPU, GPU, and memory units on a chipset. This method is being used to design Google's much awaited next gen TPUs.

 

Quotes

Quote

 

This usually takes months for a chip-design engineer to come up with, while Google’s new AI can design it in six hours’ time. The developers fed over 10,000 chip floorplans to help the AI system learn the task and come up with an optimal floorplan of a chipset that can deliver improved performance by using less power

 

My thoughts

This looks cool. My guess is that AI will lead to me getting fired in 2040, as I know batshit about AI.

 

Sources

A graph placement methodology for fast chip design | Nature

Google's AI Software Can Design a Chipset in Just Six Hours | Beebom

Google is using AI to design its next generation of AI chips more quickly than humans can - The Verge

Now Google is using AI to design chips, far faster than human engineers can do the job | ZDNet

[poochyena]

is it actually as good or better than human designs though?

[dizmo]

Not surprising. There's a pletorha of jobs that can and will be replaced by AI. Other jobs will pop up that we can't even imagine. It's going to be a tough transition for some. 

 

AI is also superb at designing material objects to be strong yet incredibly light in ways human engineers simply can't. At least not in any economical capacity. 

[Forbidden Wafer]

I know AI is awesome, but the hype makes me hate it.

Engineers spend that much time because the layout is optimized by a multitude of goals, and there are known algorithms used to help.

Beware of people saying something is "optimal", especially when dealing with AI. It usually is fundamentally wrong. It may be great/good/reasonable, but not optimal.
In this specific case, that is "easily" verifiable by comparing with deterministic optimization algorithms that do produce optimal results (not really, as multi-objective optimization = NP-Hard to verify).

Didn't look the article nor links, but I believe it were the crackhead journalists claiming something to be optimal, once again.

[PeachGr]

We suppose to use technology to work less since the beginning of time. And yeah be jobs will appear, like Youtubers and cinema actors, sport players were not a thing a century ago.

Anyway that seems like AI creates AI and that's kinda scary

[Taf the Ghost]

This feels a lot more like "easy headline" and "we couldn't hire good enough engineers to do the work" issue than some great advancement. A massive amount of chip design is already automated, and layout isn't just about the design itself. The node and objectives of the silicon will matter quite a lot. I could see this being useful after a design team (note, it's going to be an entire team) has figured out all of the requirements. That would allow the program to find extra efficiencies. 

 

This sounds a lot more like a precursor to "Google TPU delayed again". lol

 

@WolframaticAlphaAI isn't going to change as much as everyone is currently projecting. Frankly, it'll only really work well in fully digitized information sets. Further, all it's really good at is optimizing within the framework it has been given to operate and upon the information that is available. It's change some things a lot, but it'll actually put even more demand on "Systems" Engineers.

 

It's already like 3D Printing. It's really important, but it's got its place in all things, rather than taking over everything.

[Taf the Ghost]

@Forbidden WaferAI will be good for optimization runs on many designs and engineering tasks, but AI is always limited by the framework into which it is set. So it'll be good at Chess but terrible at judging Sports outcomes.

[Forbidden Wafer]

1 minute ago, Taf the Ghost said:

@Forbidden WaferAI will be good for optimization runs on many designs and engineering tasks [...]

I believe it will end up being used as an heuristic, to speed up deterministic optimization or Monte Carlo simulations.

[igormp]

Can someone give me a tldr on what's new? The AI chip design chip by google isn't new: https://www.nextplatform.com/2020/02/20/google-teaches-ai-to-play-the-game-of-chip-design/

I wonder what's actually the news here.

 

37 minutes ago, Forbidden Wafer said:

I know AI is awesome, but the hype makes me hate it.

Engineers spend that much time because the layout is optimized by a multitude of goals, and there are known algorithms used to help.

Beware of people saying something is "optimal", especially when dealing with AI. It usually is fundamentally wrong. It may be great/good/reasonable, but not optimal.
In this specific case, that is "easily" verifiable by comparing with deterministic optimization algorithms that do produce optimal results (not really, as multi-objective optimization = NP-Hard to verify).

Didn't look the article nor links, but I believe it were the crackhead journalists claiming something to be optimal, once again.

It may not be optimal, but should be good enough to fit all of the needed requirements, that's how GAs/evolutionary strategies work (and that may be the "optimal" case for a specific scenario).

 

 

 

[Forbidden Wafer]

13 minutes ago, igormp said:

It may not be optimal, but should be good enough to fit all of the needed requirements, that's how GAs/evolutionary strategies work (and that may be the "optimal" case for a specific scenario).

Good and optimal are different things.

Evolutionary strategies are often used as heuristics to speed up the search for optimal results. That's exactly where I think AI will be used.

[DrMacintosh]

Reject humans. Embrace automation. 
 

On a more serious note, I’m surprised technology like this hasn’t already been in use. I didn’t think people would be designing chips by hand in 2021. 

[igormp]

39 minutes ago, Forbidden Wafer said:

Good and optimal are different things.

Evolutionary strategies are often used as heuristics to speed up the search for optimal results. That's exactly where I think AI will be used.

Yeah, but all of the ML branch of AI ends up being just a heuristic in the end, and that's good enough for us, even if it gets stuck in some local minima.

If optimal, deterministic algorithms were available/not so costly (most of those are NP-hard as you said yourself), we would be using then instead.

[ARikozuM]

Nothing created by an AI should be patentable. 

[Bombastinator]

I am reminded that “calculator” used to be a profession before it was a thing that fit in your pocket before it was an app that fit on a thing you put in your pocket.

[Taf the Ghost]

5 hours ago, DrMacintosh said:

Reject humans. Embrace automation. 
 

On a more serious note, I’m surprised technology like this hasn’t already been in use. I didn’t think people would be designing chips by hand in 2021. 

The design phase is done on systems, as you have to design your units. However, the construction of them and then layout are automated processes. Mostly. (Obviously, each design team works a little different.)  Something about wasting 10 million USD because of a mistake you made rather than letting the automated systems do their job, comes to mind.

[cj09beira]

6 hours ago, Taf the Ghost said:

This feels a lot more like "easy headline" and "we couldn't hire good enough engineers to do the work" issue than some great advancement. A massive amount of chip design is already automated, and layout isn't just about the design itself. The node and objectives of the silicon will matter quite a lot. I could see this being useful after a design team (note, it's going to be an entire team) has figured out all of the requirements. That would allow the program to find extra efficiencies. 

 

This sounds a lot more like a precursor to "Google TPU delayed again". lol

 

@WolframaticAlphaAI isn't going to change as much as everyone is currently projecting. Frankly, it'll only really work well in fully digitized information sets. Further, all it's really good at is optimizing within the framework it has been given to operate and upon the information that is available. It's change some things a lot, but it'll actually put even more demand on "Systems" Engineers.

 

It's already like 3D Printing. It's really important, but it's got its place in all things, rather than taking over everything.

the problem with 3d printing is the materials it can print, but slowly metal printers are becoming a thing (a cheaper thing), then many more use cases appear, though if you really want to you can already use 3d printers to cast metal parts, all you need is some sand.

5 hours ago, DrMacintosh said:

Reject humans. Embrace automation. 
 

On a more serious note, I’m surprised technology like this hasn’t already been in use. I didn’t think people would be designing chips by hand in 2021. 

people make the blocks, but most of the layout is probably already done with auto-routing  software, i remember hearing some time ago that was one of intel's advantages as they did more of the layout manually so they ended up with better optimized layouts, not sure how true that is of course.

[Doobeedoo]

Very cool. 

[Master Disaster]

Step one of judgement day, AI that can create its own CPUs....

[Heliian]

Ai can create a lot of things, doesn't mean they're useful or quality.

 

[WolframaticAlpha]

3 hours ago, Heliian said:

Ai can create a lot of things, doesn't mean they're useful or quality.

 

They're supposedly faster and more efficient than their 2017 TPUs. That in my books is GOOD.

[RejZoR]

14 hours ago, poochyena said:

is it actually as good or better than human designs though?

We've had computer generated chips and they were pretty crap. Besides, chips aren't just layout.

[orbitalbuzzsaw]

14 hours ago, poochyena said:

is it actually as good or better than human designs though?

This is the question lol

[DildorTheDecent]

Rough idea as to what the design of the chip looked like:

 

 

Quote

the study explains that the system was trained on over 10,000 microchip designs, and that when faced with a selection of macro blocks to arrange in the floorplanning stage of microchip design, novelty iterations of components were found to outperform those designed by teams of human engineers.

Sauce: https://www.techpowerup.com/283235/ai-designed-microchips-now-outperform-human-designed-ones

[Bombastinator]

1 minute ago, DildorTheDecent said:

Rough idea as to what the design of the chip looked like:

 

 

Sauce: https://www.techpowerup.com/283235/ai-designed-microchips-now-outperform-human-designed-ones

This reminds me of chess and go or the whole machine compile thing.  It is possible to compile stuff by hand and there were people that claimed they could do it better. 

[sof006]

"AI" = Algorithm.

 

AI doesn't exist yet, its very annoying seeing the media misuse the term.