Friday, March 18th 2022
Google Uses Artificial Intelligence to Develop Faster and Smaller Hardware Accelerators
Designing Artificial Intelligence / Machine Learning hardware accelerators takes effort from hardware engineers in conjunction with scientists working in the AI/ML area itself. A few years ago, we started seeing AI incorporated into parts of electronic design automation (EDA) software tools, helping chip designers speed up the process of creating hardware. What we were "used to" seeing AI do are just a couple of things like placement and routing. And having that automated is a huge deal. However, it looks like the power of AI for chip design is not going to stop there. Researchers at Google and UC Berkeley have made a research project that helps AI design and develop AI-tailored accelerators smaller and faster than anything humans made.
In the published paper, researchers present PRIME - a framework that created AI processors based on a database of blueprints. The PRIME framework feeds off an offline database containing accelerator designs and their corresponding performance metrics (e.g., latency, power, etc.) to design next-generation hardware accelerators. According to Google, PRIME can do so without further hardware simulation and has processors ready for use. As per the paper, PRIME improves performance upon state-of-the-art simulation-driven methods by as much as 1.2x-1.5x. It also reduces the required total simulation time by 93% and 99%, respectively. The framework is also capable of architecting accelerators for unseen applications.
As far as comparing the test results, PRIME was in its prime time (pun intended). Google's very own EdgeTPU was compared to PRIME-made design, and the AI-generated chip was faster with latency improvement of 1.85x, resulting in a much quicker design. Researchers also noticed a fascinating thing, and that the framework-generated architectures are smaller, resulting in a more minor, less power-hungry chip. You can read more about PRIME in this free-to-access paper here.
Sources:
Google AI Blog, via The Register
In the published paper, researchers present PRIME - a framework that created AI processors based on a database of blueprints. The PRIME framework feeds off an offline database containing accelerator designs and their corresponding performance metrics (e.g., latency, power, etc.) to design next-generation hardware accelerators. According to Google, PRIME can do so without further hardware simulation and has processors ready for use. As per the paper, PRIME improves performance upon state-of-the-art simulation-driven methods by as much as 1.2x-1.5x. It also reduces the required total simulation time by 93% and 99%, respectively. The framework is also capable of architecting accelerators for unseen applications.
49 Comments on Google Uses Artificial Intelligence to Develop Faster and Smaller Hardware Accelerators
This is a specialised AI designed to shine in one task , not many.
General AI that's some dreamy shit IMHO.
Your computer this is not recognized by Google, open your smartphone ... you will receive special code .... enter that to unlock your account.....
Did this once ... three days later again .... Your computer this is not recognized by Google, open your smartphone ... you will receive special code .... enter that to unlock your account.....
Do they have any physical address? because I do plan to mail them few farts with 100% natural ingredients.
I do wonder what sort of optimisations could be achieved if this was applied to x86, which - being old as the hills - has a proportionally much larger percentage of human-hand-tuned design that could potentially benefit massively.
Autobans for racism, xenophobia, intolerance, trolling, propoganda, and botting are something Google sorely needs some AI accelerators to help with; Maybe it's just me but I find 90% of all Youtube videos to be marred somewhere by the worst examples that humanity has to offer.
Every time that I do log-in, from then and since, its actually a new benchmark, if my account this is still alive.
Google does not have any definition of it ideology.
And because of that, I can not take them seriously, and neither I care of what they might do, with the few KB of data which this is called as user profile.
The wise internet user, he should avoid all internet services which violating his soul, then his logic, and also avoid everything which seems as low importance.
Google this dreams having me tracked, and be a force of control of my habits, and of my thoughts, I have bad news for them.
I am autonomous H.I. and this will never change, for them I am an super dangerous terrorist. :)
Either way, its one runs his business according to his own business plan.
And regarding business plans, I do prefer my own.
No doubt technology and techniques have advanced leaps and bounds since then, but it's not 'a new thing'.
Trick is to have the optimum from high-level design to low-level trace routing done automatically. And trust me, no human can do it better than AI at this point.
The issue that persists is still humans are more betterer (I mean that) at figuring out random things than hardware is with a limited sample set. A chip may perform great with X and Y but not Z and Z may be the manufacture node, the few subsets of computation. Considering AMD still has cache latency issues and Intel can’t make a competitive chip without sacrificing branching security issues tells a lot about the status of everything, Arm/Apple has to conflate numbers and still ended up with a massive cache system to reach their performance need.
I said, quite unambiguously that "90% of youtube videos are marred", not that 90% of youtube commenters should be autobanned.
The difference between those two things is multiple orders of magnitude. 90% of those videos have at least one example of racism, xenophobia, intolerance, trolling, propoganda, or botting - all of which are ToS violations punishable by account ban according to Google.
At any rate, 9/10 videos there's either an argument with bigots or between bigots risen to the first page of comments, or it's spammed by bots/adverts/scammers somewhere in the comments chain.
I can't think of any AI algorithm with the power of a dog's sense of smell.
A dog's sense of smell operates on the entire world: a very large data set.
Dogs can do many things. Most AI routines can only do one thing.
Dogs are pretty good at training themselves to be dogs, better then any AI routine.
Wolfram has some good lectures on the subject:
www.wolfram.com/wolfram-u/catalog/machine-learning/
My own experience with AI routines is in time series analysis.
They give interesting results but are generally inferior to a well thought out adaptive filter.
If you know of a publicly available classification algorithm that is designed for martingale time series please post link.
I would be happy to have my opinion changed.
...and no, I would not let a dog tell me the distribution of a time series.
Your comment is one of the clearest demonstrations of how no matter how good someone / some Company is, there will always be "better" people who will bring them down.
I wonder how do you think you succeeded in anything, considering the way you think Google failed so much...
As for Google, thank you for all the good you have bring to the world and to my life: the Internet was reborn the day I saw the Google search start page. YouTube is better and better: there is so much good content available (e.g. high quality technical tutorials) that the problem is having time to see the top 1% of them. And although I use other email services for certain purposes (e.g. ProtonMail), nothing compares with Gmail, both in the desktop and in Android.:) exactly