New Laser-Array Processor Could Improve AI Computing Efficiency (viterbischool.usc.edu)

Researchers from USC and MIT published a Nature Photonics study describing a laser-array optical processor that could dramatically cut the energy and space costs of AI compute. The team demonstrated a photonic chip with more than 400 surface-emitting lasers (VCSELs) packed into 1 cm² that directly converts electronic memory into light at a 10 GHz clock — equivalent to 10 billion neural activations per second. They report up to a 100× improvement in both throughput and energy efficiency over current ML processors, with the per-conversion energy down to several attojoules (roughly the energy of ten visible photons), which is 5–6 orders of magnitude better than modern optical modulators. The authors say near-term engineering could yield another two orders of magnitude gain, implying far larger improvements ahead. This matters because large DNNs (training GPT-4 used ~25,000 GPUs and ~50 MWh of electricity) are hitting electronic and thermal bottlenecks. Previous optical neural networks suffered from poor electro‑optic conversion, large device footprints, crosstalk, and latency from lacking inline nonlinearity; integrating dense laser arrays addresses many of those limits by increasing compute density and lowering conversion energy. If scalable, laser-array optoelectronic processors could accelerate ML workloads from data centers to edge devices, cutting carbon and enabling new architectures for training and inference.