Photonic Chip Performs Image Recognition at the Speed of Light
spectrum.ieee.org
spectrum.ieee.org
> do the computation faster than anyone can stream the data into the chip. The more interesting problem is data movement, as usual.
Is that not the selling point of integrated photonic circuits? You got photonic chips/components that are connected by photonic waveguides. So essentially photons replace electrons for all intents and purposes and the data moves at close to the speed of light.
1. Photonics is a new technology. It is more expensive than silicon 2. Silicon can do the compute as fast as photonics, and we understand silicon. Building a silicon chip is 'easy'. 3. The problem is data transport, which silicon cannot do fast enough. We can compute inferences faster than we can get the data
Photonic compute doesn't help with (3). Photonic data transport does, and there are several companies who have entered this market, including several who have already exited with large deals. However, photonic data transport can terminate at photonic compute or silicon compute, and silicon compute is fast enough and cheaper / better understood.
These are not exactly high res images.
That doesn’t limit the creation of new models and new understanding. Just because scaling might be different doesn’t mean it’s impossible.
Our existing models did not exist at some point in the past.
When we say that a CPU is running at some frequency that's essentially the speed at which a signal propagates through all the gates of its slowest pipeline stage, and the individual gates have a transition frequency an order of magnitude higher.
So if a CPU is running at 4GHz, it completes a clock cycle in 0.25ns (250ps), but that means that individual transistors must have latencies <=0.025ns (25ps). If you built a tiny little 9-neuron single-layer network out of those transistors then the network's latency would be around the single-gate latency and it would appear to be 20x better than this photonic setup, and even if those 9 neurons were laid out sequentially it would outperform this photonic chip by 2x.
Now granted, that's comparing the performance of a mature IC process using equipment that cost billions of dollars to set up, to something that was built in a university lab on a substantially smaller budget, so direct performance comparisons aren't exactly fair in that sense. But it's also misleading to take a direct hardware implementation of an itty bitty neural network and handwave away the very real issues with scaling such an approach.
As an aside, I recently learned of https://lightmatter.co/, a photonic chip manufacturer
Veritasium has two great videos on it:
https://www.youtube.com/watch?v=GVsUOuSjvcg
https://www.youtube.com/watch?v=IgF3OX8nT0w
I found a hackernews discussion for the second one: https://news.ycombinator.com/item?id=29645610