» Current | 2018 | 2017 | 2016 | Subscribe

Linley Newsletter

Graphcore Makes Big AI Splash

September 18, 2018

Author: Linley Gwennap

Graphcore’s much anticipated first product is a powerful processor that aims to loosen Nvidia’s firm grip on deep-neural-network (DNN) training in large data centers. Starting from a blank page, the British company created a new AI-specific architecture, which it calls the IPU (intelligence processing unit), that packs more than a thousand programmable cores onto a single massive chip. The result is a highly flexible design that generates as many floating-point operations (flops) as Nvidia’s V100 at less than half the power. Graphcore claims its performance advantage will be greater across a broad range of neural networks thanks to the IPU’s large on-chip memory and other unique features. The first IPU chip, called the GC2, is already sampling and is scheduled for production by the end of this year.

At $112 million, Graphcore is one of the best-funded AI-chip startups. Sequoia, a top venture-capital firm, has invested $50 million; strategic investors include Bosch, Dell, and Samsung. Graphcore has 116 employees and two dogs, most at its Bristol, U.K., headquarters plus a handful at its rapidly expanding Silicon Valley office.

The GC2 supports FP32 and FP16 data formats, similar to Nvidia’s V100, making it well suited to DNN training. It can also perform inferencing, but most inference tasks are shifting to more-efficient integer formats that run well on CPUs and ASICs such as Google’s TPU1. Whereas GPUs rely on a host CPU to handle certain DNN layers, the IPU can handle all layers, increasing efficiency by avoiding data transfers to the host. Graphcore will sell GC2-based accelerator cards to Dell and other system vendors; these cards will generate up to 250Tflop/s at a 300W TDP—twice the performance of a V100 card at a similar power. The startup offers ONNX and TensorFlow drivers, easing the implementation of DNNs on its chip.

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

Events

Linley Fall Processor Conference 2018
Covers processors and IP cores used in embedded, communications, automotive, IoT, and server designs.
October 31 - November 1, 2018
Hyatt Regency, Santa Clara, CA
More Events »

Newsletter

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