Reacción a "Photonic chips developed to improve computing speed and reduce power consumption"

What do you think of the studies?

‘Both are of very good quality, although they address different aspects of photonic computing. The Harris et al paper is more complex as it describes a complete solution that co-integrates photonics and electronics using Global Foundries processes. It is a 128x128 processor in silicon photonics. It actually describes Lightmatter's ENVISE product. The focus of this article is on describing the complete development of this processor and conducting experiments and obtaining results with a view to comparing it with the performance obtained with 32-bit neural network-based electronic computing solutions. The development includes a very detailed description of the electronics (rather than the photonics) and the results are very promising, although not yet fully competitive in some AI tasks ( transformer models). It is a rather difficult article to read because it requires knowledge of many areas (electronics, photonics and software).

Peng's article focuses more on the photonics part of the processors and how they can be scaled to be able to perform matrix-vector calculations with minimal latency in the optical segment. He also describes a development of a silicon matrix (in this case, 64x64) processor and its application to heuristic (i.e., ad hoc) problem solving. In this case, the aim is to emulate Ising machines. The value of this work lies mainly in having been able to scale the photonic chip to include 16,000 components while guaranteeing low latency (delay) processing’.

How do they fit with existing evidence and what new developments do they bring?

‘Both studies represent an advance over the state of the art. The Harris processor is the first complete solution to co-integrate high-density photonics and electronics and RISC V processors. The result is a compact board with very substantial processing power. Although it still lags behind electronic solutions, some of the results obtained put it very close to them.

If the technology could be scaled up, the impact would be significant, as it could compete with and even surpass purely electronic solutions. It is precisely a way forward in scaling up that is proposed in Peng's article and is therefore also very important’.

Are there any limitations to be taken into account?

‘The most important limitations lie in the scaling of the solutions (including more photonic components to produce matrix-vector processors with more inputs), the limitation in optical input power to avoid non-linearities in silicon photonics chips to improve computational performance (especially in the Harris processor), and the management of noise associated with the inherently analogue computation process.