Taiwan-based Artilux, a company specializing in GeSi photonics technology and CMOS-based SWIR sensing, announced a new AI photonics computing platform called Inception.
Artilux says that its new Inception platform, based on novel hybrid optoelectronics, delivers orders-of-magnitude improvements in both power and area efficiencies compared to current standard computing platforms, without relying on advanced CMOS process nodes and active cooling.
At the core of the Artilux Inception is an original optoelectronics-enabled hybrid systolic array architecture compatible with existing CPU/GPU/TPU/LPU, enabling large-scale general matrix–matrix multiplication (GEMM) — the dominant workload in modern AI computing — with no data skew and negligible propagation delay. The hybrid analog/digital nature of this architecture enables massive parallel dot-product computation in a single step using a dense 2D array of optoelectronic neurons (OENs). Each OEN consists of a light emitter, a GaN microLED, coupled with a photodetector (GeSi pixel), and an in-pixel memory.
Artilux says that its new platform can be built using mature CMOS process nodes and integration technologies that are compatible with existing semiconductor manufacturing infrastructure. Artilux also announced that its first-generation processing core based on Artilux Inception is now under development
