» Current | 2021 | 2020 | 2019 | Subscribe

Linley Newsletter

Lightmatter Shines in Data-Center AI

March 30, 2021

Author: Mike Demler

Lightmatter aims to deliver the industry’s first photonic AI processor. After successfully testing the Mars prototype it introduced last year, the startup is scaling the architecture by 16x for its Envise design. On the basis of test results and cycle-accurate simulations running Bert-Base inference, it expects Envise to provide about six times better performance per watt than Nvidia’s DGX A100.

Like the Mars prototype, Envise is a 3D chip stack. The bottom layer combines two photonics-based tensor-processor chips, each integrating a 128x256 (32K) ALU array. The ALUs execute multiply-accumulate (MAC) operations, as well as pooling and similar linear-algebra functions, using photons rather than electrons. The chips employ the same 90nm GlobalFoundries technology as the prototype. An interposer connects the bottom layer to a 12nm ASIC that controls neural-network execution, computes nonlinear functions such as ReLU activations, handles general-purpose compute tasks, and provides the converters for the analog photonic ALUs.

In typical operating conditions, the entire accelerator consumes 80W. The company plans to ship Envise in a 4U server that includes 16 of the devices, along with 1TB of DDR4 memory and two 200Gbps smart network interface cards (NICs), consuming 3.1kW (TDP). It also intends to develop a 36U rack system that contains nine Envise servers and a top-of-rack switch.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »

Events

Linley Spring Processor Conference 2021
April 19 - 23, 2021
Virtual Event
More Events »