Graphcore, a UK-based company that develops accelerators for AI workloads, introduced the second generation of its Intelligence Processing Units (IPUs) this morning, which will soon be available on the company’s M2000 IPU machine. Graphcore claims that this new GC200 chip will enable the M2000 to achieve a petaflop of computing power in a package that measures the width and length of a pizza box.
AI accelerators like the GC200 are a type of specialized hardware that was developed to accelerate AI applications, especially artificial neural networks, deep learning and machine learning. They are often multi-core in design and focus on low-precision arithmetic or in-memory computing. Both can increase the performance of large AI algorithms and lead to state-of-the-art results in natural language processing, computer vision, etc. and other domains.
The M2000 is powered by four of the new 7-nanometer GC200 chips, each of which combines 1
In addition to the M2000, according to Graphcore, customers can connect up to 64,000 GC200 chips for up to 16 exaflops of computing power and petabytes of memory and support AI models with theoretically trillions of parameters. This is made possible by Graphcore’s IP fabric connection technology, which supports data transmission with low latency up to a speed of 2.8 Tbit / s and establishes a direct connection to IPU-based systems (or via Ethernet switches).
The GC200 and M2000 work with Graphcore’s bespoke poplar, a graphics toolchain optimized for AI and machine learning. It can be integrated into Google’s TensorFlow framework and Open Neural Network Exchange (an ecosystem for interchangeable AI models) and in the latter case offers a full training period. Preliminary compatibility with Facebook’s PyTorch was introduced in the fourth quarter of 2019. Full functional support followed in early 2020. The latest version of Poplar – version 1.2 – introduced replacement memory management features to take advantage of the GC200’s unique hardware and architecture design in terms of memory and memory data access.
Graphcore, founded by Simon Knowles and Nigel Toon in 2016, has so far raised over $ 450 million from Robert Bosch Venture Capital, Samsung, Dell Technologies Capital, BMW, Microsoft and AI. Arm co-founder Hermann Hauser and DeepMind co-founder Demis Hassabis valued at $ 1.95 billion. The first commercial product was a 16-nanometer PCI Express card – C2 – that was available in 2018. This package was launched on Microsoft Azure in November 2019. (Microsoft also uses Graphcore products internally for various AI initiatives.)
Earlier this year, Graphcore, in collaboration with Dell, announced the availability of the DSS8440 IPU server and launched Cirrascale IPU-Bare Metal Cloud, an IPU-based managed service offering from cloud provider Cirrascale. More recently, Graphcore has announced some of its early customers – including Citadel Securities, Carmot Capital, Oxford University, JP Morgan, the Lawrence Berkeley National Laboratory, and European search engine company Qwant – and open source solutions for creating and creating GitHub libraries Run apps on IPUs.
Graphcore may have momentum on its side, but it has competition in a market that is expected to reach $ 91.18 billion by 2025. In March, Hailo, a startup that develops hardware to marginally accelerate AI inferences, raised $ 60 million in venture capital. California-based Mythic raised $ 85.2 million to develop its own in-memory architecture. Mountain View-based Flex Logix launched an inference coprocessor in April that claims to deliver ten times the throughput of existing silicon. Last November, Esperanto Technologies secured $ 58 million for its 7-nanometer AI chip technology.