They claim that "the photonics core operates at the speed of light (...) boosting bandwidth by a factor of ten while reducing latency from the typical 100ns with electronics-based chips to a staggering 100 picoseconds (a 1000X improvement)". Plus a radical reduction in power consumption ...
It would be interesting to hear from anyone with insights into the realistic potential of this tech.
The matrix-multiplication is partially done in an "analog" way, multiplication with constant coefficients can be done using programmed optical attenuation elements, and the summing stage can be done by superposition of multiple optical signals.
This works well if you can do with a low dynamic range (low precision math) and with seldomly-changing coefficients (i.e. multiplying a changing vector/matrix with a constant matrix). On the other hand you save orders of magnitude of energy per operation. Maybe for neural networks this is a favourable trade-off. The paper at the link above talks about a 4-bit precision matrix multiplication.
The problem of these approaches is that they are limited by the electronics needed to control the optics. It is all about the electronics. The numbers presented for the chip were worse than what can be accomplished with conventional digital electronics at equivalent bit depth.
Could you put this interferometer array directly after a lens? Before the integration readout happening on a digital camera sensor. You wouldn't have the photon-electron-photon conversion.
At least some computation can already be done before the integration. There's already some feature highlighting happening on the retina.
I assume it's not possible, you need a laser as input source.
7 comments
[ 2.9 ms ] story [ 23.0 ms ] threadIt would be interesting to hear from anyone with insights into the realistic potential of this tech.
https://aip.scitation.org/doi/10.1063/5.0001942
which seems to describe a similar technology.
The matrix-multiplication is partially done in an "analog" way, multiplication with constant coefficients can be done using programmed optical attenuation elements, and the summing stage can be done by superposition of multiple optical signals.
This works well if you can do with a low dynamic range (low precision math) and with seldomly-changing coefficients (i.e. multiplying a changing vector/matrix with a constant matrix). On the other hand you save orders of magnitude of energy per operation. Maybe for neural networks this is a favourable trade-off. The paper at the link above talks about a 4-bit precision matrix multiplication.
At least some computation can already be done before the integration. There's already some feature highlighting happening on the retina.
I assume it's not possible, you need a laser as input source.