China's AI Analog Chip Claimed To Be 3000X Faster Than Nvidia's A100 GPU

tom's HARDWARE ^

[FarCenter]

A new paper from Tsinghua University, China, describes the development and operation of an ultra-fast and highly efficient AI processing chip specialized in computer vision tasks. The All-analog Chip Combining Electronic and Light Computing (ACCEL), as the chip is called, leverages photonic and analog computing in a specialized architecture that’s capable of delivering over 3,000 times the performance of an Nvidia A100 at an energy consumption that’s four million times lower. Yes, it’s a specialized chip – but instead of seeing it as market fragmentation, we can see it as another step towards the future of heterogeneous computing, where semiconductors are increasingly designed to fit a specific need rather than in a “catch-all” configuration.

As published in Nature, ACCEL is quoted as hitting 4.6 trillion operations per second in vision tasks – hence the 3,0000x performance improvement against Nvidia’s A100 (Ampere) and its 0.312 quadrillion operations. According to the research paper, ACCEL can perform 74.8 quadrillion operations per second at 1 W of power (what the researchers call “systemic energy efficiency) and a computing speed of 4.6 peta-operations per second. Nvidia’s A100 has since been superseded by Hopper and its 80-billion transistors H100 super-chip, but even that looks unimpressive against these results.

Of course, speed is essential in any processing system. However, accuracy is necessary for computer vision tasks. After all, the range of applications and ways these systems are used to govern our lives and civilization is wide: it stretches from the wearable devices market (perhaps in XR scenarios) through autonomous driving, industrial inspections, and other image detection and recognition systems in general, such as facial recognition. Tsinghua University’s paper says that ACCEL was experimentally tried against Fashion-MNIST, 3-class ImageNet classification, and time-lapse video recognition tasks with “competitively high” accuracy levels (at 85.5%, 82.0%, and 92.6%, respectively) while showing superior system robustness in low-light conditions (0.14 fJ μm−2 each frame).

[DUMBGRUNT]

It is occasionaly said, there are lies, damn lies, and benchmarks.
Sorry to Mark T.

That said, Tom’s Hardware is a reputable source.

[TexasGator]

Since it has discrete operations why is it called analog?

[MNDude]

I used AI to explain what this means:
Hey, check this out! There’s a new chip from Tsinghua University in China that’s crazy fast for AI stuff. It’s called ACCEL, and it’s all about computer vision tasks. Apparently, it’s like 3,000 times faster than Nvidia’s A100 GPU, and it uses way less energy.

So, this chip does a whopping 4.6 trillion operations per second for vision tasks, which is way better than what the A100 can do. And get this, it only needs 1 watt of power to do all that. That’s some serious energy efficiency!

They tested it on things like image recognition and video stuff, and it did really well. Like, it got 85.5% accuracy for Fashion-MNIST, 82.0% for 3-class ImageNet classification, and a crazy 92.6% for time-lapse video recognition. Plus, it works great in low-light conditions.

Speed is awesome, but accuracy is super important for things like self-driving cars and facial recognition. So, this chip could be a game-changer for all sorts of cool tech.

And yeah, I know it’s called an “analog” chip, but it still does digital stuff. Tech names can be weird sometimes! 🚀🤖

[BereanBrain]

I am a 30+ year HPC expert. We will go electrical (analog) and optical for Neural Nets. Our brain is analog and works on chemical signaling which is 1,000,000 times (roughly) slower than digital circuits. Yet we are pretty good at vision.

[ridesthemiles]

RIGHT-——AND MY 1979 BUICK HAS 1500 HORSEPOWER

[VAFreedom]

So they stole someone else’s time machine design and went to the future to steal Nvidia’s M1 GPU?