Graphcore, a member of the Experimental Infrastructure for Exploration of Exascale Computing (eX3), has announced details of its next generation IPU-POD systems, or Machine Intelligence compute at scale. This system is specifically engineered to meet the future compute requirements of machine intelligence by tightly coupling large numbers of Graphcore Intelligence Processing Units (IPUs), to provide a massively parallel environment for heavy AI workloads. This technology is already available in eX3.
“We committed early on to Graphcore technology, acquiring one of the first Graphcore IPU-POD64 systems, counting 64 IPU processors with over 77,000 cores and 8 PetaFLOPS mixed precision AI compute. Right now, we are diving into this new technology to fully explore its power,” said Professor Are Magnus Bruaset, research director at Simula Research Laboratory. “As part of the national research infrastructure eX3 funded by the Research Council of Norway, our IPU-POD will be available to HPC researchers throughout the country.”
Large machine intelligence workloads require compute performance far beyond what is possible by scaling up existing hardware; a new system approach is needed. Graphcore systems, including its IPU-PODs are elastic, and can scale up and scale out to the level of compute needed to meet future machine intelligence requirements. This is a game changer – a new kind of hardware that allows for innovative breakthroughs in machine intelligence, and in a fraction of the time.
Through eX3, Graphcore and Simula have initiated a close research collaboration expected to yield results that will be important to both science and business.“eX3 and Simula are really the perfect partners for Graphcore. They are looking ahead of the curve, and recognize the emerging challenges that will define next generation Machine Learning compute, such as large models and growing sparsity,"said OlaTørudbakken, SVP Systems, Graphcore. "Most of all, we are delighted that they are helping to provide our technology to some of Europe’s leading innovators, ensuring that our continent is able to help define the future of AI.”
Innovations are sure to come, as researchers are already finding new ways to use the IPU-POD in eX3. “It will surely be used for its original purpose in computing large AI/ML workloads,” said Professor Bruaset. “However, at Simula we also aim to push the envelope of this architecture by investigating how it can be applied to other types of sparse computations, such as in the algorithmic kernels of physics-based simulations.”
This topic is also at the center of their first joint venture “SparCity”. This EuroHPC project, in which Simula and Graphcore are partners with research institutions in Turkey, Germany, and Portugal, was recently granted after receiving a perfect score (15/15) from the evaluation of this competitive call. Next, Simula and Graphcore, together with the other eX3 partners, will this fall pursue funding of eX3-II, an extension of the eX3 infrastructure until 2026.
The eX3 project
The eX3 project (Experimental Infrastructure for Exploration of Exascale Computing) is currently funded for five years (2018-2022) through the Research Council of Norway’s program for national research infrastructures.
The eX3 infrastructure is not meant to be an exascale computer by itself; rather, it will be a carefully curated ecosystem of technology components that will be crucial for embracing exascale computing. eX3 will allow HPC researchers throughout Norway and their collaborators abroad to experiment hands-on with emerging HPC technologies – hardware as well as software. These researchers will work closely with owners of computationally challenging problems that require physics-based and/or data-driven modelling to be solved.
The eX3 project is hosted by Simula, with the following partners: Graphcore AS, the University of Bergen, University of Tromsø, Norwegian University of Science and Technology (NTNU), Uninett Sigma2 AS, Dolphin Interconnect Solutions AS, Numascale AS, Fabriscale Technologies AS, and Skala Technologies AS.
Graphcore is a UK-based semiconductor company founded in 2016, specifically addressing the computational challenges of machine intelligence through the tailored design of a new type of processor, the Intelligence Processing Unit (IPU). Graphcore’s technology center in Oslo has been fundamental in developing the IPU-POD; the elastic and scalable multiprocessor system that will bring the IPU technology to the HPC world.
- eX3 website: ex3.simula.no
- Apply for access to eX3: application information and form
- Graphcore press release: https://www.graphcore.ai/posts/introducing-second-generation-ipu-systems-for-ai-at-scale