Photons instead of electrons: IBM offers new paths for processors and AI

Modern computing capabilities of classical processor architectures have exhausted themselves, IBM is confident. Moreover, they have become an obstacle to the development of machine learning and artificial intelligence systems. The breakthrough is seen in the development of silicon photonics and in-memory computing, where data is processed where it is stored. And today IBM has proven that they have groped the way to the electronics of the future, in which photons will fly through circuits instead of electrons.

IBM specialists, together with scientists from several countries, have developed and implemented an optical computing system to accelerate the operation of neural networks. In particular, the company has created a "photonic tensor core", which is capable of performing the so-called convolution operation - a mathematical operation on two functions that displays the third function - in a one-time step. This is usually a simple addition or multiplication, but processing one piece of data requires billions of such operations, so low latency and low consumption are vital requirements for such systems.

Performing operations on data in memory is an additional opportunity to save both on consumption and on latency since data does not need to be transferred to the processor and back. In IBM's development, data was stored and processed in phase-change memory locations.

The next step in speeding up data processing is Wavelength Division Multiplexing (WDM). To put it simply, the data entered the memory unit in the form of light with different wavelengths. This approach allows both to expand the data transmission channel (frequency extension) and to carry out operations on the photon data stream in parallel. Where electrons flowed in series in chains, photonic chains allow for parallel data flow and simultaneous processing of each of the flows. This is a colossal speedup of data processing!

As an experiment, a 9 × 4 matrix was created with a maximum of four input vectors per time step, each of which was transmitted as light with its own wavelength. For MAC (multiply-accumulate) operations, the matrix performed 2 TOPS / s at a modulation rate of 14 GHz. IBM expects the proposed circuitry to achieve in-memory performance of photonic circuits of PetaMAC / s per mm² (thousands of trillion MAC operations), three orders of magnitude higher than current 1 TOPS / mm² for current electronics.

Source: ibm.com