Graphcore, the British semiconductor company that develops accelerators for AI and machine learning, has greatly increased the performance of its massively parallel Intelligence Processing Unit (IPU) servers.
Graphcore sells its AI-oriented chips in rack-mounted designs called IPU-PODs. Up until now, the maximum per cabinet has been 64 units (in the IPU-POD64). The newest racks are twice and four times as large: the IPU-POD128 and IPU-POD256.
With 32 petaFLOPS of AI compute in the IPU-POD 128 and 64 petaFLOPS in the IPU-POD 256, Graphcore says it can now extend its reach into AI supercomputer territory, and with a much smaller footprint than the typical supercomputer, which could fill a basketball court.
The IPU-POD systems disaggregate the AI compute from the servers, which means different types of AI workloads requiring different levels of performance can all be run on the same POD. For example, a POD can be allocated for faster training of large Transformer-based language models across an entire system, or the system can be divided into smaller, flexible vPODs to give more developers IPU access.
This is done through Graphcore’s Poplar software stack, which includes an SDK, the Graphcore Communication Library (GCL) for managing communication and synchronization between IPUs, and PopRun and PopDist, which allow developers to run their applications across multiple IPU-POD systems.
For intra-rack IPU communication, earlier PODs used 64Gb/s IPU-Links. Now, the IPU-POD128 and IPU-POD256 use new Gateway Links, a horizontal, rack-to-rack connection that extends IPU-Links using tunneling over regular 100Gb Ethernet.
Both systems have been developed for cloud hyperscalers, national scientific computing labs and enterprise companies with large AI teams in markets like financial services or pharmaceutical. Graphcore’s initial customers include the U.S. Department of Energy's Lawrence Berkeley National Laboratory and J.P. Morgan.
IPU-POD16, IPU-POD64, IPU-POD128 and IPU-POD256 are shipping to customers today from French IT giant ATOS and other systems integrator partners around the world and are available to buy in the cloud from Cirrascale.
Copyright © 2021 IDG Communications, Inc.