News China's ACCEL Analog Chip Promises to Outpace Industry Best in AI Acceleration for Vision Tasks

[P.Amini]

I am learning

 

[Amdlova]

LAST time I see someone telling about ten years we will have flying cars... so propaganda for everyone...

 

[neojack]

I dont really understand the article, but when i read the title i was like 'fake news..." then i saw it was published in Nature

So is it a technological breakthrough ? what are the real life usages of such kind of chip ?

 

[derekullo]

I can only assume that when they asked for an Nvidia A100 chip to compare performance to some one forged an A100 tag onto a Geforce 2 MX

 

[COLGeek]

Please stay on topic and leave politics to other sites. Thank you.

 

[TCA_ChinChin]

The Nvidia A100 isn't really a GPU either (it doesn't have a display output)

Thats true, but my points still stands. They are vastly different computing systems that don't really have a lot of good comparison between them. I'm just saying its a little irrelevant that the Chinese university is claiming its so much faster than the A100.

Its like claiming an RTX4090 is better at playing Cyberpunk compared to a quatum computer. It's probably true, but it doesn't really mean anything.

 

[changetasm]

My vision task is seeing a few problems with the math here.

This.

My vision task is seeing a few problems with the math here.

4.6 trillion is less than 0.312 quadrillion. Tom's Hardware article error.

 

[George³]

This.

4.6 trillion is less than 0.312 quadrillion. Tom's Hardware article error.

As far as I find articles with an earlier date and similar text, and even almost the same title, I don't know if the material here can be considered copyrighted. If it is a reprint, the colleague who posted it may not have the right to make changes to the content, even though it contains mathematical inaccuracies.

 

[bit_user]

The Nvidia A100 isn't really a GPU either (it doesn't have a display output)

More than that, it's virtually devoid of TMUs, ROPs, and other graphics-specific hardware engines. I still call it a "GPU", in certain contexts, because it's a lot easier than the verbal contortions like "GPU-like compute accelerator", which I also sometimes say.

 

[bit_user]

So is it a technological breakthrough ? what are the real life usages of such kind of chip ?

Cheap, low-powered computer vision. I already gave an example of a battery-powered doggy door that will only open for your dog. IMO, that's the end-goal of such tech. It doesn't need to be very high-resolution - all it needs to be is cheap and low-power.

Other examples: embedding gesture recognition in low-cost electronics, like lamps and coffee makers. I'm not saying these are great examples, but they give you an idea of what sorts of things would be enabled that isn't feasible today.

 

[zsydeepsky]

Cheap, low-powered computer vision. I already gave an example of a battery-powered doggy door that will only open for your dog. IMO, that's the end-goal of such tech. It doesn't need to be very high-resolution - all it needs to be is cheap and low-power.

Other examples: embedding gesture recognition in low-cost electronics, like lamps and coffee makers. I'm not saying these are great examples, but they give you an idea of what sorts of things would be enabled that isn't feasible today.

or products that are very sensitive to energy capacity. A good example is Drone.
DJI drones have many cameras so they can avoid collisions in the air, and running multiple cameras for vision recognition becomes power hungry for what a drone could offer and creates overheating issues. this tech will basically eliminate the problem.
another example is VR/AR headsets, they need to do vision recognition & 3D environment SLAM, and all of them eat away power & performance capacity like a troll.

 

[dlheliski]

A new research paper in Nature published by Tsinghua University describes a new analog computing chip that beats Nvidia's A100 at 3,000 times its performance and 4,000 million times higher energy efficiency at computer vision tasks. Considering the potential markets for devices such as these and the potential for further miniaturization within some aspects of ACCEL's architecture, scaling might want to find a place for its fabrication.

China's ACCEL Analog Chip Promises to Outpace Industry Best in AI Acceleration for Vision Tasks : Read more

This article is so riddled with errors, it is complete gibberish. They claim 4550 TOPs, not 4.6. And the .312 quadrillion ops of Nvidia is 312 TOPS so it is only *15* times faster, not 3000. However, the paper compares to some unknown statemtent that Nvidia is 156 TOPs which does substantiate a 30X ratio. However both of these are dishonest since Nvidia can achieve 1248 TOPs in int8, which is closer to the precision of this. Therefore it is really about 4X faster. 4X, not 3000.

The power numbers are also gibberish. They claim 7.48e4 TOPs/W vs. 0.52 TOPS/w for nvidia. This is about 15,000 not the billions you claim. They use their own measurement that produces .001 TOPs/w which is unclear why it is 500X lower than Nvidia, but even this is only a 7.5e7 power ratio.

Did you even read the abstract? It only claims 3 orders of magnitude better power and 1 order of magnitude better speed, but you are claiming nearly 13 orders of magnitude better power and 3 orders magnitude better speed. So you are exaggerating by factors of 10,000,000,000 and 100 respectively.

 

[George³]

This article is so riddled with errors, it is complete gibberish. They claim 4550 TOPs, not 4.6. And the .312 quadrillion ops of Nvidia is 312 TOPS so it is only *15* times faster, not 3000. However, the paper compares to some unknown statemtent that Nvidia is 156 TOPs which does substantiate a 30X ratio. However both of these are dishonest since Nvidia can achieve 1248 TOPs in int8, which is closer to the precision of this. Therefore it is really about 4X faster. 4X, not 3000.

The power numbers are also gibberish. They claim 7.48e4 TOPs/W vs. 0.52 TOPS/w for nvidia. This is about 15,000 not the billions you claim. They use their own measurement that produces .001 TOPs/w which is unclear why it is 500X lower than Nvidia, but even this is only a 7.5e7 power ratio.

Did you even read the abstract? It only claims 3 orders of magnitude better power and 1 order of magnitude better speed, but you are claiming nearly 13 orders of magnitude better power and 3 orders magnitude better speed. So you are exaggerating by factors of 10,000,000,000 and 100 respectively.

WoW, sir!
Okay, now, there are indeed some numbers in the article that may be incorrect. However, I also read your numbers and am also shocked. Perhaps we all need a consultation with a mathematician who is in good health right now.

 

[dlheliski]

WoW, sir!
Okay, now, there are indeed some numbers in the article that may be incorrect. However, I also read your numbers and am also shocked. Perhaps we all need a consultation with a mathematician who is in good health right now.

I'm a retired PhD in Elec Engineering, if that helps.

 

[PaulAlcorn]

Note: We unfortunately made a mistake in the performance calculations in the article and have updated it to correct those issues.

 

[George³]

I'm a retired PhD in Elec Engineering, if that helps.

Yes, it definitely helps. Could you, if you have enough time and curiosity, scour the original publication in Nature or in scientific archives and find the correct performance values? By the way what is TOP's? If I'm not mistaken, 1TOPs is a trillion operations and some errors in article here may come from that. It is also important to clarify what precision the numbers refer to. This has already been asked in the comments, but is missing an answer.
Edit, in actualised article now have all info that was answered in the topic. Thanks!

 

[vanadiel007]

South Korean kitchen for superconductor. Bad memory, sir.

Ah yes, the kitchen sink around the corner of the Chinese kitchen sink.
I took a left and should have taken a right.

I am sure this chip will be smashing and on par with all the other Chinese chips.

 

[vanadiel007]

Note: We unfortunately made a mistake in the performance calculations in the article and have updated it to correct those issues.

4,000 million did not give the error away? lol
Can't wait for the next super conductor update to be honest. Will be an interesting read.

 

[usertests]

What's an order of magnitude or three between friends?

 

[sitehostplus]

Wait, I thought we were baking a super conductor in the Chinese kitchen sink last time I visited.
Good thing it's 4,000 million billion times more energy efficient. I mean, 4,000 million. Errr, 4 billion, 0.004 trillion? umm, very big much better!

That caught my eye too. Unless my math is off, would that not indicate the device in question is using somewhere in the neighborhood of nanowatts of power? Or even less?

 

[bit_user]

or products that are very sensitive to energy capacity. A good example is Drone.
DJI drones have many cameras so they can avoid collisions in the air, and running multiple cameras for vision recognition becomes power hungry for what a drone could offer and creates overheating issues. this tech will basically eliminate the problem.
another example is VR/AR headsets, they need to do vision recognition & 3D environment SLAM, and all of them eat away power & performance capacity like a troll.

Good examples. The only reason I'm reluctant to embrace them is that I don't know how dependent they are on resolution. For something like VR/AR SLAM, I think you need probably about 1024x1024 and I have no idea if their approach would be cost-effective to produce at that resolution. For a drone, you might need to go even higher.

 

[George³]

resolution 1024x1024

I think is enough. Maybe more wide?

 

[bit_user]

I think is enough. Maybe more wide?

Sure, for object classification at such close range. That's why I thought a doggie door or a home appliance were good examples.

Cyberdyne Systems Model 101 (T-800) notwithstanding, I'm not sure it'd be sufficient for a general-purpose robot's main vision system.

 

[Chung Leong]

If AI is so darn advanced, why am I still cleaning stupid spam from my mailbox everyday?

 

[bit_user]

If AI is so darn advanced, why am I still cleaning stupid spam from my mailbox everyday?

Dunno. My GMail has a lot less spam than other services I've used. At work, we use Office 365 and it also stays pretty free of junk.

I think it ultimately comes down to three things:

1. How many resources the email service is willing to devote to the problem. To the extent they use AI models, they would need to be many orders of magnitude cheaper than something like ChatGPT, due to the amount of mail that needs to be scanned.
2. Not everyone would regard the same email as spam. I mean, there are spammy spams, but then there are also commercial & political emails that some would want to see while others wouldn't.
3. Your tolerance for false-positives vs. false-negatives (i.e. legit email getting caught in the spam filter vs. spam reaching your inbox). In general, people have a lower tolerance for false-positives than false-negatives. So, that would tend to weight spam classifiers towards being a little more conservative.