Hacker News new | past | comments | ask | show | jobs | submit login
New chip opens door to AI computing at light speed (phys.org)
38 points by thunderbong 10 months ago | hide | past | favorite | 10 comments



The Nature Photonics article can be found here:

https://www.nature.com/articles/s41566-024-01394-2

This appears to be the same paper published on arxiv:

https://arxiv.org/abs/2302.00793

The abstract describes the actual problem that these photonic chips solve:

> We designed and experimentally demonstrated a vector–matrix product for a 2 × 2 matrix and a 3 × 3 matrix. We also designed a 10 × 10 matrix using the proposed 2D computational method. These examples demonstrate that these techniques have the potential to enable larger-scale wave-based analogue computing platforms.

It appears the main contribution of this paper is not so much a new chip technology, but a new way to efficiently model the flow of light through a particular kind of chip (using raised and lowered bumps of silicon) that makes designing photonic chips that compute matrix multiplications practical.

For a fixed matrix, they can design a chip that approximately multiplies that matrix by an input query vector, using light.

Being a fixed matrix severely limits the usefulness of this for now, especially at the 2x2 and 3x3 sizes, but it could be useful if it could be made much larger, or ways to vary the contents of the matrix are discovered.


Thanks for the summary

This won't work alone for ML. The matrix sizes are huge with billions of parameters. Doing light based calculations for that seems impossible with this technique.

Potentially you could break apart a big matrix into smaller matrices for these chips to process


A billion parameter matrix, if it was square, is 32,000 by 32,000. NVidia's tensor units are basically 4x4 matrix multipliers.


“The potential for large scale” I think was the operative phrase. A PoC at 2x2, 10x10 etc demonstrates scaling characteristics and potential.


I've been reading about how the chips that use light are 'just around the corner' since the 1990's, much like the mythical battery technologies and other fabled tech beasts. Have there been any tangible products so far? It seems it has never been outside labs.


> "virtually unhackable" technology due to computations occurring simultaneously without the need for storing information in working memory

Why would that have anything to with vulnerabilities and hacking?




The title makes it sound general, but this seems to be only for a single baked-in specific neural net.


this seems pretty incredible if true.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: