A brand new technical paper titled “Scalable Coherent Optical Crossbar Structure utilizing PCM for AI Acceleration” was revealed by researchers at College of Washington.

“Optical computing has been not too long ago proposed as a brand new compute paradigm to fulfill the calls for of future AI/ML workloads in datacenters and supercomputers. Nevertheless, proposed implementations to date endure from lack of scalability, giant footprints and excessive energy consumption, and incomplete system-level architectures to turn out to be built-in inside current datacenter structure for real-world purposes. On this work, we current a very scalable optical AI accelerator primarily based on a crossbar structure. Now we have thought-about all main roadblocks and handle them on this design. Weights can be saved on chip utilizing part change materials (PCM) that may be monolithically built-in in silicon photonic processes. All electro-optical parts and circuit blocks are modeled primarily based on measured efficiency metrics in a 45nm monolithic silicon photonic course of, which could be co-packaged with superior CPU/GPUs and HBM reminiscences. We additionally current a system-level modeling and evaluation of our chip’s efficiency for the Resnet-50V1.5, contemplating all vital parameters, together with reminiscence measurement, array measurement, photonic losses, and vitality consumption of peripheral electronics. Each on-chip SRAM and off-chip DRAM vitality overheads have been thought-about on this modeling. We moreover handle how utilizing a dual-core crossbar design can eradicate programming time overhead at sensible SRAM block sizes and batch sizes. Our outcomes present {that a} 128 x 128 proposed structure can obtain inference per second (IPS) just like Nvidia A100 GPU at 15.four occasions decrease energy and seven.24 occasions decrease space.”

Discover the technical paper here. Printed October 2022.

arXiv:2210.10851v1. Authors: Daniel Sturm, Sajjad Moazeni.

Source link


Please enter your comment!
Please enter your name here