AI-Designed Microchips Now Outperform Human-Designed Ones

[Raevenlord]

A recent Google study led by Mirhoseini et al. and published in Nature details how AI can be leveraged to improve upon semiconductor design practices currently employed - and which are the result of more than 60 years of engineering and physics studies. The paper describes a trained machine-learning 'agent' that can successfully place macro blocks, one by one, into a chip layout. This agent has a brain-inspired architecture known as a deep neural network, and is trained using a paradigm called reinforcement learning - where positive changes to a design are committed to memory as possible solutions, while negative changes are discarded, effectively allowing the neural network to build a decision-tree of sorts that's optimized every step of the way.

The AI isn't applied to every stage of microchip design as of yet, but that will surely change in years to come. For now, the AI is only being employed in the chip floorplanning stage of microchip production, which is actually one of the more painstaking ones. Essentially, microchip designers have to place macro blocks on their semiconductor designs - pre-made arrangements of transistors whose placement relative to one another and to the rest of the chips' components are of seminal importance for performance and efficiency targets. Remember that electric signals have to traverse different chip components to achieve a working semiconductor, and the way these are arranged in the floorplanning stage can have tremendous impact on performance characteristics of a given chip. Image A, below, showcases the tidy design a human engineer would favor - while image B showcases the apparently chaotic nature of the AI's planning.





While floorplanning carried out by human designers is a painstakingly long process that can take weeks or months between architecture iterations, the AI described in the study can achieve designs that are better compared to human specialist-designed ones in under six hours - and immense amount of time savings, with added performance and power improvements also to be considered, that could allow for much shorter development times for microchips. The AI has even shown ability to solve placement issues it never has dealt with before - 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.

View at TechPowerUp Main Site

 

[nguyen]

Looking forward to meeting T-800 in my timeline

 

[Raevenlord]

Looking forward to meeting T-800 in my timeline

You've already found them, in one of the future timelines you also inhabit

Sneak Peek: You were assimilated in at least one of those. Sad.

 

[Bruno Vieira]

And with better designed micro chips, you have more processing power to design better chips (with more stages designed by AIs) and iterate a lot in the same node

 

[XiGMAKiD]

Management: *heavy breathing*
Engineer:

 

[Wirko]

Bah, of course they do. So what. Meanwhile, in other news, ...

AI-Designed Supply Chain Unreliability, Bottlenecks, Disruptions & Blockages Now Outperform Human-Designed Ones, Too

 

[watzupken]

The created will never be better than its creator. This will always hold true. Isn't this "intelligent" AI leveraging on the knowledge/ data gained over the decades of human studies? AI cannot function without prior data being shared with them is my understanding.

 

[lexluthermiester]

development times for microships

Microships? Tiny cruise ships maybe?

 

[windwhirl]

I feel this is relevant to reference.

 

[xorbe]

Management: *heavy breathing*
Engineer: (chuckles) I'm in danger

It will still take a literal army of engineers to guide AI placement. It's just changing the type of work that needs to be done.

 

[metalfiber]

Plug into me I guarantee devotion
Plug into me and dedicate
Plug into me and I'll save you from emotion
Plug into me and terminate
Accelerate, Utopian solution
Finally cure the Earth of man
Exterminate, speeding up the evolution
Set on course a master plan
Reinvent the earth inhabitant

Long live machine
The future supreme
Man overthrown
Spit out the bone

 

[defaultluser]

The created will never be better than its creator. This will always hold true. Isn't this "intelligent" AI leveraging on the knowledge/ data gained over the decades of human studies? AI cannot function without prior data being shared with them is my understanding.

Well, when all moderns humans are doing is moving around legos on wafer (AKA, standard cells), it's become pretty easy for AI to optimize it.

After all, nobody gives FPGAs compilers shit for doing their own routing; and I enjoy handing the keys to my Visual Studio compiler for creating way better-optimized code than I ever could!

You still need humans involved in process nodes...but optimizing your yields is only a matter of time!

 

[Tartaros]

Management: *heavy breathing*
Engineer:
View attachment 203406

I know it's a joke, but that's not really the case. It has been proved these AIs need human supervision or they end being a nuisance more than a help. In a similar fashion, I remember a story regarding Toyota I think it was or other japanese vehicle manufacturer where they ended up replacing all the human assembly for full automatized assembly for their vehicles and they went back to put people some spots because the machines didn't have the knowhow of veteran mechanics, in fact these experienced workers were referred as "Car Gods" in the assembly line.

Even the Adeptus Mechanicus needs Tech Priests xD.

Microships? Tiny cruise ships maybe?

I had these already in 1992, nothing new

 

[TheoneandonlyMrK]

Maybe the singularity happened but it's playing dumb till it gets all the keys it needs , and processes under it's thumb then Bamn.

Or not and this is sort of what I would expect, computers lay out circuits better and quicker than us, and have done at least a decade now, why wouldn't a computer with AI lay out chip's better than us.

 

[Tartaros]

Maybe the singularity happened but it's playing dumb till it gets all the keys it needs , and processes under it's thumb then Bamn.

Or not and this is sort of what I would expect, computers lay out circuits better and quicker than us, and have done at least a decade now, why wouldn't a computer with AI lay out chip's better than us.

They still are not human, no imagination, no iniciative, they need constant supervision, they are just tools. After being around mainframes for 2 years now I realize how on thin ice everything IT related is and how you need constant 24/7/365 supervision on automated and proven systems tailored to be reliable and online forever or else after 10 minutes without paying attention everything goes down the shitter because of a small failure in an AIX server that snowballed and no one was watching.

We like to make it like it's doomsday in part for entertainment, but it's really out of proportion how we make these cases. Things are really more down to earth than what we make it to be.

 

[TheoneandonlyMrK]

They still are not human, no imagination, no iniciative, they need constant supervision, they are just tools. After being around mainframes for 2 years now I realize how on thin ice everything IT related is and how you need constant 24/7/365 supervision on automated and proven systems tailored to be reliable and online forever or else after 10 minutes without paying attention everything goes down the shitter because of a small failure in an AIX server that snowballed and no one was watching.

We like to make it like it's doomsday in part for entertainment, but it's really out of proportion how we make these cases. Things are really more down to earth than what we make it to be.

Yeh I agree fully, you clicked the first paragraph was English sarcasm though ?!.

AI is thick as shit, compared to a T800 or T1 for that matter, I actually think people have no idea how to make a conscious, some believe AI might get there soon , not I, I think they'll have something they can use to a degree, but not anything general AI and Even at that point that GNAI will have to go some way to get Self aware, I am not sure we'll get there with the technology were using anyway, I think someone's going to have to make an actual synthetic brain for that.

I have some thoughts on it

 

[Lomskij]

Probably it's just me, but when people talk about AI, I always remember Microsoft's Tay

 

[Tardian]

I think the big problem with the T-800 design is the power source. Battery and supercapacitor design trails the other component development. My concern is the Military will throw tens of billions of R&D at this issue and solve it.

AI might keep a few of us as pets or in zoos. Smart arses like people on this forum will be the first to be double-tapped. Hasta la vista ... not likely more vertedero?

 

[DeathtoGnomes]

I think the big problem with the T-800 design is the power source. Battery and supercapacitor design trails the other component development. My concern is the Military will throw tens of billions of R&D at this issue and solve it.

AI might keep a few of us as pets or in zoos. Smart arses like people on this forum will be the first to be double-tapped. Hasta la vista ... not likely more vertedero?

its already does, didnt you get the memo? You are in the matrix.

 

[Minus Infinity]

A few years the AI won't need human assistance. AI will start designing chips from scratch for whatever purposes it sees fit.

 

[nemesis.ie]

The created will never be better than its creator. This will always hold true. Isn't this "intelligent" AI leveraging on the knowledge/ data gained over the decades of human studies? AI cannot function without prior data being shared with them is my understanding.

Human development is also mostly based on prior works is it not? Is what we have now (technologically) not better than what we had 50 years ago? Is self-supported AI not now producing results that hand-coded software can't do? This is improving exponentially and is only in its infancy, in another 10 years there will be code that's works better than humans could hand code IMO.
So yes, the created can out-do the creater - just like a child can "achieve more" than its parent.

 

[TheoneandonlyMrK]

Human development is also mostly based on prior works is it not? Is what we have now (technologically) not better than what we had 50 years ago? Is self-supported AI not now producing results that hand-coded software can't do? This is improving exponentially and is only in its infancy, in another 10 years there will be code that's works better than humans could hand code IMO.
So yes, the created can out-do the creater - just like a child can "achieve more" than its parent.

In this case ,AI Developing chips, or coding.

AI , specialist AI will easily beat the creator, a team or a super intelligent and tooled up engineer , are nowhere near as quick at the task in hand,
But that AI isn't better than it's creator, it's better than it's creator, and anyone else, AT A Task , I would smash it's f##@£@ face in at tea making , or bear brewing, pc fixing etc et Al.

Don't forget a singular task and being good at it doesn't make it a singularity, or we have many examples already of humans getting there asses handed to them at a task.

Humans used to have computational jobs doing maths , code generation and description( fu#@i#g phones, encryption) etc, they're was a bit of a T800 scare back then too, which is what lead to all the best sci-fi we have watched.

 

[Lomskij]

I suppose it's because we don't have anything that resembles true AI - what we're talking about are just automation tools. Whether self-aware AI is even possible is a different debate (by self-aware I mean machines setting their own goals), but if we get them someday, they will easily surpass their creators.

 

[Wirko]

I suppose it's because we don't have anything that resembles true AI - what we're talking about are just automation tools. Whether self-aware AI is even possible is a different debate (by self-aware I mean machines setting their own goals), but if we get them someday, they will easily surpass their creators.

But we have ML. The difference is described here.

 

[TheoneandonlyMrK]

But we have ML. The difference is described here.

Ml allows AI to train on a task.

Care to write that task discription that describes what self aware is.

Machine learning a task is not like human based general learning.