Research team demonstrates world’s fastest optical neuromorphic processor | Hacker News Reader

Research team demonstrates world’s fastest optical neuromorphic processor

swinburne.edu.au

103 pointssizzle20 comments

ArsTechnica article discussing this paper (and one other), with a bit of background information: https://arstechnica.com/science/2021/01/two-ways-of-performi...

There are several universities, big corporations, and startups taking another look at analog computing, and ML seems to be application that will make it viable.

Instinctively ML feels like it would mesh well with the continuous distributions in analog systems better than digital.

Just simply pause for a second and admire this part: 10 trillion per second. No matter how you look at it, that's a lot.

That's a lot, but the way to look at it is: "how many TOPS per second per watt when we run Resnet-50 on it"

The ars article linked above¹ says that one advantage that optical processors may have is that the majority of the heat generated by the system comes from the laser source, not the processor itself.

Imagine 'turning on' a datacenter by flipping the switch on a 10MW laser :D

¹ https://news.ycombinator.com/item?id=25691582

Aren't there a ton of tiny lasers in the processor itself?

No; it’s one laser source turned into a frequency comb by the magic of saturable absorbers and Fabry-Perot resonators.

Edit: sorry, looks like micro-ring resonators in this paper. These are amazing structures.

I’m not knowledgeable on this topic. But I always dreamt of “optical” processor that does OR AND XOR functions. And I thought it was impossible. Because you can’t do this with photons.

So when this article claims that optical CPU of the neuromorphic type has been reached, were photon based CPUs always real and practical?

You can do even better, namely analog computation.

Yes! Some fascinating research happening as we speak with in-memory ML using ReRAM.

Always nice to see optical processors in the news.

How do the 10 TeraOPs/s on this compare to the 11 on the Apple M1?

It's far too early to compare this with any commercial NN accelerators. From the article:

Although the performance of ONNs is not yet competitive with leading-edge electronic processors at >200 TOPS (for example, Google TPU and other chips), there are straightforward approaches towards increasing our performance both in scale and speed. Further, with a single processor speed of 11.3 TOPS, our VCA is approaching this range. The CA is fundamentally limited in data size only by the electrical digital-to-analogue converter memory, and processing 4K-resolution (4,096×2,160 pixels) images at >7,000 frames per second is possible. The 720 synapses of the CNN (72 wavelengths per synapses per neuron, 10 neurons), a substantial increase for optical networks, enabled us to classify the MNIST dataset. Nonetheless, further scaling is needed to increase the theoretical prediction accuracy from 90% to that of state-of-the-art electronics, typically substantially greater than 95%.

7k fps on 4k resolution is an absurd amount of bandwidth. For 8 bit pixels that is 619 Gbps. Very interesting. Thanks.

For comparison, Nvidia A100 has 2TB/s memory bandwidth. That's 16k Gbps.

Typically when you mention a frame rate in the context of a processor performance, you care about throughput, not input data transfer bandwidth. Max data transfer bandwidth can limit max throughput, but usually not in the first layer. For example, a standard benchmark is Resnet-50 Imagenet frame rate.

absurd bandwidth? try this on for size: the square kilometer array does online (“realtime” loses its meaning when the data processing latency gets long enough) interferometry on millions of antenna achieving an OUTPUT data rate of a few TB/s, which is then filtered down to a paltry 1 TB/s by an exaflop computer. it’s a mind blowing instrument, and as a public science project there is lots of fun stuff to read about it :)

Apropos nothing, your username has taught me a new word of Dutch.

^_^

If it doesn't work yet on MNIST it's too early to talk about 4K images being processed at 7000/s.

"10x the performance of TPU"

arxiv paper: https://arxiv.org/pdf/2011.07393.pdf

Reply on news.ycombinator.com