First step to self-awareness - Google AI designs its own chip

[williamcll]

Google's own Artificial intelligence has managed to create it's own Tensor processor after less than a day of training.

Quote

Ideally you want a chip that’s optimized to do today’s AI, not the AI of two to five years ago. Google’s solution: have an AI design the AI chip. “We believe that it is AI itself that will provide the means to shorten the chip design cycle, creating a symbiotic relationship between hardware and AI, with each fueling advances in the other,” they write in a paper describing the work that posted today to Arxiv. “We have already seen that there are algorithms or neural network architectures that… don’t perform as well on existing generations of accelerators, because the accelerators were designed like two years ago, and back then these neural nets didn't exist,” says Azalia Mirhoseini, a senior research scientist at Google. “If we reduce the design cycle, we can bridge the gap.”

 

Mirhoseini and senior software engineer Anna Goldie have come up with a neural network that learn to do a particularly time-consuming part of design called placement. After studying chip designs long enough, it can produce a design for a Google Tensor Processing Unit in less than 24 hours that beats several weeks-worth of design effort by human experts in terms of power, performance, and area. Placement is so complex and time-consuming because it involves placing blocks of logic and memory or clusters of those blocks called macros in such a way that power and performance are maximized and the area of the chip is minimized. Heightening the challenge is the requirement that all this happen while at the same time obeying rules about the density of interconnects. Goldie and Mirhoseini targeted chip placement, because even with today’s advanced tools, it takes a human expert weeks of iteration to produce an acceptable design.

 

Goldie and Mirhoseini modeled chip placement as a reinforcement learning problem. Reinforcement learning systems, unlike typical deep learning, do not train on a large set of labeled data. Instead, they learn by doing, adjusting the parameters in their networks according to a reward signal when they succeed. In this case, the reward was a proxy measure of a combination of power reduction, performance improvement, and area reduction. As a result, the placement-bot becomes better at its task the more designs it does.

The team hopes AI systems like theirs will lead to the design of “more chips in the same time period, and also chips that run faster, use less power, cost less to build, and use less area,” says Goldie.

Source: https://spectrum.ieee.org/tech-talk/semiconductors/design/google-invents-ai-that-learns-a-key-part-of-chip-design

https://arxiv.org/abs/2003.08445

Thoughts: I saw this coming long ago, it might be useful in discovering new routing for data and I could see it saving a lot of time in some scenarios. I don't have a use for NPUs myself so I do wish this could be applied to more consumer processors.

[Guest]

Up next, it designs its own batteries.

[TechyBen]

37 minutes ago, williamcll said:

Google's own Statistical analysis system has managed to create it's own Tensor processor after less than a day of maths.

Source: https://spectrum.ieee.org/tech-talk/semiconductors/design/google-invents-ai-that-learns-a-key-part-of-chip-design

https://arxiv.org/abs/2003.08445

Thoughts:

Fixed it.

[CarlBar]

8 minutes ago, huilun02 said:

So Google is doing the 'chicken or egg first' to themselves

 

In this case, Chicken.

 

That makes me hungry, are abstract code concepts edible i wonder?

[RejZoR]

AMD's Bulldozer chips were also computer generated weren't they? It was faster and cheaper, but we all know how that ended...

[givingtnt]

Oooooo

[matrix07012]

7 hours ago, Den-Fi said:

Up next, it designs its own batteries.

Humans in Matrix ARE the processors, the final script says they are an energy source because the studio feared the audience wouldn't understand it.

[Guest]

46 minutes ago, matrix07012 said:

Humans in Matrix ARE the processors, the final script says they are an energy source because the studio feared the audience wouldn't understand it.

I'm just gonna go with the version of the script that works with the semi-joke I made, lol.

But I do appreciate the insight.

[Taf the Ghost]

Considering most SoCs are into the billions of transistor range, almost all of it is done by algos designed by the companies already. Google basically is claiming they've managed to make a new algo to improve workflow. 

 

Somehow, we're in 2020 and it feels like 1995 with "and this mundane thing happened on THE INTERNET!" when it comes to "AI". All it is using a neural network to optimize already existing algorithms. Process designed to optimized pre-existing processes optimized All the Things.

[Curufinwe_wins]

2 hours ago, Taf the Ghost said:

Considering most SoCs are into the billions of transistor range, almost all of it is done by algos designed by the companies already. Google basically is claiming they've managed to make a new algo to improve workflow. 

 

Somehow, we're in 2020 and it feels like 1995 with "and this mundane thing happened on THE INTERNET!" when it comes to "AI". All it is using a neural network to optimize already existing algorithms. Process designed to optimized pre-existing processes optimized All the Things.

Sounds like the shit my current employer tries to say when they claim AI is making all these optimizations when it's nothing more than a simple neural network being constrained by human contextual boundaries (don't get me wrong, I expect Google's effort is way better than ours, but still ugh)

[GOTSpectrum]

 

That is all!

[Taf the Ghost]

3 hours ago, Curufinwe_wins said:

Sounds like the shit my current employer tries to say when they claim AI is making all these optimizations when it's nothing more than a simple neural network being constrained by human contextual boundaries (don't get me wrong, I expect Google's effort is way better than ours, but still ugh)

For as hard as actual underlying technology is to comprehend to most, the reality is that all a neural network system can never do is optimize to the criteria inputted and the assumptions made when getting there. The results are always "surface-level" and very "numb", but they are optimized. In the case of placing transistors (which is one of the few places in the world of Human Experience where billions of iterations are actually needed), it's actually quite useful to optimize well over that many iterations. The real power of automation has always been to iterate at a rate beyond what a human can do.

 

That being said, AI is a massive, overblown bubble. Spreadsheets & Databases have been extremely powerful since being introduced, but it still takes a lot of skill to get anything useful out of them. AI will be even worse. Unless you have a task that can be iterated on millions of times like SoC development, the actual utility will always be very limited.

[RejZoR]

4 hours ago, comander said:

Intel's Nehalem and Sandy Bridge parts were as well... Same with Zen and Zen2

Basically everyones' chips use algorithms to lay out critical parts. 

https://en.wikichip.org/wiki/File:amd_zen_core.png <- look at this core. The "mushroomy" stuff was designed mostly by an AI. 

What the humans are doing is providing high level guidance, taking the AI's design, tuning it and then getting it to work in the real world.

As someone who writes "AI" for a living, there's A LOT of tuning to do. 

So, it's not really "Ai", it's just layout coordinator that stuffs required stuff into as small space possible. "Ai". Just throw it at anything. I guess I'm eating an Ai sardines right now...

[Taf the Ghost]

1 hour ago, RejZoR said:

So, it's not really "Ai", it's just layout coordinator that stuffs required stuff into as small space possible. "Ai". Just throw it at anything. I guess I'm eating an Ai sardines right now...

They seem to have put in the work to optimize another step that used to require a lot more of "hands-on" work. If their criteria was set properly. It'll still need to be reviewed by the engineers that would have needed to do weeks of silicon implementation work. So, it might, eventually, be a solid time saver. (Still took nearly a day on what is likely a near super computer.)

 

Humorously, to me, it looks like they used the data they got from OpenAI playing Dota 2 for how to attack the problem. Also quite funny is I think OpenAI took ages longer to get good, lol.

[Taf the Ghost]

7 hours ago, VegetableStu said:

"strange, the traces look like a game's battle map..."

I wouldn't doubt they pick all of the Cheesy Meta heroes as well.

 

The actual OpenAI Dota 2 project was interesting, mostly as because after millions of hours of playing Dota 2, the AI basically just ran at you and all it did was out-execute a human team. Its Borg-like unrelentingness is what allowed it to do so well.

[Curufinwe_wins]

32 minutes ago, Taf the Ghost said:

I wouldn't doubt they pick all of the Cheesy Meta heroes as well.

 

The actual OpenAI Dota 2 project was interesting, mostly as because after millions of hours of playing Dota 2, the AI basically just ran at you and all it did was out-execute a human team. Its Borg-like unrelentingness is what allowed it to do so well.

Which is why the SC2 results were WAY more impressive. Even after the additional limitations placed on Alphastar. It plays pretty darn cool legit games. It can be tricked, eventually, sometimes, but there are actual lessons being learned that have already been incorporated into proplay.

[Taf the Ghost]

6 minutes ago, Curufinwe_wins said:

Which is why the SC2 results were WAY more impressive. Even after the additional limitations placed on Alphastar. It plays pretty darn cool legit games. It can be tricked, eventually, sometimes, but there are actual lessons being learned that have already been incorporated into proplay.

The AlphaStar stuff was definitely interesting, though the fact that the AI could keep a recallable database of builds that the opponent could do, it made scouting and the AI's ability to execute always have a pretty sizable advantage.  Though, SC2 is a little weird in that there were already unbeatable AIs by I think 2011, since you can run at over 30k APM within the game engine. 

[Curufinwe_wins]

10 minutes ago, Taf the Ghost said:

The AlphaStar stuff was definitely interesting, though the fact that the AI could keep a recallable database of builds that the opponent could do, it made scouting and the AI's ability to execute always have a pretty sizable advantage.  Though, SC2 is a little weird in that there were already unbeatable AIs by I think 2011, since you can run at over 30k APM within the game engine. 

I don't think I'd agree with those assessments on the whole, in fact one of Alphastars apparent weaknesses (while still quite good) was to just really bizarre play. (The one map TLO won was by basically forcing Alphastar into a cyclical paralysis position until he was able to bring overwhelming force to bear)

 

Plus the later variants capped APM both spike and average to levels well below the best players and even incorporated limits to the speed and frequency by which screen could jump around (after all otherwise the fact that a computer could theoretically take in a screens worth of information almost instantly would be a signficant advantage).

 

Though yes, between games of perfect information, and unlimited control, there were hard to beat AIs in the past (they tended to be fairly trivial to exploit, but if you just let them get to where their strength was... good luck)

 

 

Anyways, not really relevant this discussion. If you are interested more, BeastyQt has some great commentary/analysis of later alphastar games on his youtube channel. Not just for experts of the game or such.

[TechyBen]

38 minutes ago, Curufinwe_wins said:

I don't think I'd agree with those assessments on the whole, in fact one of Alphastars apparent weaknesses (while still quite good) was to just really bizarre play. (The one map TLO won was by basically forcing Alphastar into a cyclical paralysis position until he was able to bring overwhelming force to bear)

 

Plus the later variants capped APM both spike and average to levels well below the best players and even incorporated limits to the speed and frequency by which screen could jump around (after all otherwise the fact that a computer could theoretically take in a screens worth of information almost instantly would be a signficant advantage).

 

Though yes, between games of perfect information, and unlimited control, there were hard to beat AIs in the past (they tended to be fairly trivial to exploit, but if you just let them get to where their strength was... good luck)

 

 

Anyways, not really relevant this discussion. If you are interested more, BeastyQt has some great commentary/analysis of later alphastar games on his youtube channel. Not just for experts of the game or such.

Yeah. You can always make a bigger hammer, and a big hammer always hits hard. Difference being, a person can often work around and find new exploits to said hammer (in this example, the hammer is AI ).

Some AI try, through trial and error, to overcome such obstacles. But often perform worse than a human in intelligence, but better in number of iterations (a server farm of compute will give millions of iterations in the time a person can think of a few responses). It's again, not *really* "better intelligence" as the real life counter would be a city block of people working on a problem (example would be NASA getting the astrunoughts back in Apollo 13).

 

So really, when someone says "this AI is better" what they mean is "we spent more money/we used more resources/we have more people/we used more math/we have more compute/we outsourced".

[Kamina]

I didn't know Terminator was based on a true story that takes place into our future.

[TechyBen]

On 4/3/2020 at 11:28 PM, comander said:

While throwing compute at problems definitely matters, it's not all that. 

HOW something is computed often matters more than the quantity of compute tossed at a problem (Quick-sort on a 1000 number list is MUCH faster than BOGO sort on the same list with 10,000 the IOPs)

In a lot of cases the exact methods used matter a TON. Here and today, if pressed I could make an image classifier in a month for some sort of problem using my desktop (let's assume all of my relevant data is already gathered). A decade ago, it would've taken a team of PhDs with super computers a lot longer to come up with worse results. In many instances, people end up standing on the shoulder's of giants and reusing techniques (or pretrained neural networks). The people at DeepMind are in some sense those giants. They're REALLY on the cutting edge.

Yep, but a lot of AI implementations are BOGO. Just a *huge* set of compute thrown at the problem until the BOGO gets the results.

Though saying that, I think it's always been the case. It's expensive but easy to scale up to a large data set/compute. It's harder to think clever and use a small compute to do a lot of work.

[TechyBen]

20 minutes ago, comander said:

An aside - BOGO sort is equivalent to throwing a stack of cards in the air and checking to see if they landed in the right order. If you fail, then try again. It's the least efficient algorithm that isn't intentionally designed to be awful. 

There's more to it than that. 
15 years ago if you threw a bunch of compute at a problem using the latest and greatest algorithms... you'd get numerical instability and you'd get worthless results. 

A lot of what Google/DeepMind does is coming up with new algorithms and new tricks to implement things so that you can get the same or better results with less compute. 

For perspective, a lot of voice recognition is now being done locally on Google devices (phone, tablets, laptops, smart speakers, etc.) that used to be done on the cloud. It's faster and better than what was being done years ago with WAY more compute. That's better algorithms. Think 10,000x speed up with better results and energy efficiency. 

Yeah, I know what BOGO is.

AI can sometimes be used in the same way, just pilling data and hoping something comes out. Other times it's done well and intelligently.

 

Voice was always done locally pre-Google. I had Win 95 with voice access (only a demo though ) at one time. It was put on the cloud, as you say, to get more data to make better software. Technically "AI" helped with this. But not in the way most think (it was for the data set).