Graphcore raises $222 million to scale up AI chip production

Graphcore, a Bristol, U.K.-dependent startup creating chips and methods to accelerate AI workloads, nowadays introduced it has raised $222 million in a series E funding spherical led by the Ontario Teachers’ Pension Program Board. The investment decision, which values the corporation at $2.77 billion write-up-funds and provides its total raised to date to $710 million, will be made use of to assistance ongoing global enlargement and even further accelerate potential silicon, systems, and computer software growth, a spokesperson advised VentureBeat.

The AI accelerators Graphcore is producing — which the enterprise phone calls Intelligence Processing Units (IPUs) — are a form of specialised components made to velocity up AI purposes, especially neural networks, deep finding out, and machine learning. They are multicore in style and design and concentrate on minimal-precision arithmetic or in-memory computing, the two of which can raise the effectiveness of large AI algorithms and direct to condition-of-the-artwork effects in purely natural language processing, pc vision, and other domains.

Graphcore, which was established in 2016 by Simon Knowles and Nigel Toon, launched its initially commercial solution in a 16-nanometer PCI Categorical card — C2 — that became available in 2018. It’s this package deal that released on Microsoft Azure in November 2019 for consumers “focused on pushing the boundaries of [natural language processing]” and “developing new breakthroughs in equipment intelligence.” Microsoft is also applying Graphcore’s products internally for several AI initiatives.

Before this calendar year, Graphcore announced the availability of the DSS8440 IPU Server, in partnership with Dell, and released Cirrascale IPU-Bare Steel Cloud, an IPU-dependent managed services giving from cloud supplier Cirrascale. Much more not too long ago, the firm disclosed some of its other early consumers — between them Citadel Securities, Carmot Funds, the College of Oxford, J.P. Morgan, Lawrence Berkeley Countrywide Laboratory, and European lookup engine corporation Qwant — and open up-sourced its libraries on GitHub for building and executing apps on IPUs.

In July, Graphcore unveiled the next technology of its IPUs, which will soon be created available in the company’s M2000 IPU Device. (Graphcore states its M2000 IPU products are now shipping and delivery in “production volume” to shoppers.) The enterprise promises this new GC200 chip will empower the M2000 to attain a petaflop of processing electric power in a 1U datacenter blade enclosure that actions the width and length of a pizza box.

The M2000 is run by four of the new 7-nanometer GC200 chips, every single of which packs 1,472 processor cores (operating 8,832 threads) and 59.4 billion transistors on a one die, and it provides more than 8 moments the processing effectiveness of Graphcore’s current IPU merchandise. In benchmark tests, the corporation promises the four-GC200 M2000 ran an image classification product — Google’s EfficientNet B4 with 88 million parameters — extra than 32 situations a lot quicker than an Nvidia V100-primarily based program and over 16 periods quicker than the latest 7-nanometer graphics card. A solitary GC200 can provide up to 250 TFLOPS, or 1 trillion floating-position-operations for each second.

Outside of the M2000, Graphcore suggests prospects will be ready to link as many as 64,000 GC200 chips for up to 16 exaflops of computing ability and petabytes of memory, supporting AI types with theoretically trillions of parameters. That is manufactured achievable by Graphcore’s IPU-POD and IP-Material interconnection technology, which supports minimal-latency information transfers up to prices of 2.8Tbps and specifically connects with IPU-based mostly methods (or through Ethernet switches).

The GC200 and M2000 are created to operate with Graphcore’s bespoke Poplar, a graph toolchain optimized for AI and equipment finding out. It integrates with Google’s TensorFlow framework and the Open up Neural Community Exchange (an ecosystem for interchangeable AI products), in the latter case offering a total training runtime. Preliminary compatibility with Facebook’s PyTorch arrived in Q4 2019, with full element support pursuing in early 2020. The latest edition of Poplar introduced exchange memory management functions meant to choose advantage of the GC200’s unique components and architectural design with regard to memory and facts access.

Graphcore may well have momentum on its side, but it has competitors in a market place that is predicted to reach $91.18 billion by 2025. In March, Hailo, a startup building components made to speed up AI inferencing at the edge, nabbed $60 million in enterprise funds. California-based Mythic has elevated $85.2 million to develop customized in-memory architecture. Mountain See-primarily based Flex Logix in April launched an inference coprocessor it promises provides up to 10 periods the throughput of existing silicon. And final November, Esperanto Systems secured $58 million for its 7-nanometer AI chip technology.

Further than the Ontario Teachers’ Pension System Board, Graphcore’s collection E noticed participation from cash managed by Fidelity Global and Schroders. They joined current backers Baillie Gifford, Draper Esprit, and others.