Whether it is in energy, mobility, material sciences, medicine or environmental research – to understand natural and technical processes in all their complexity, researchers use the fastest high-performance computers in the world. As a member of the National High Performance Computing Alliance (NHR), the National High-Performance Computing Center NHR@KIT at the Karlsruhe Institute of Technology (KIT) offers supercomputing resources and a wide range of supporting services to researchers from all over the Germany.
Methods and algorithms from the areas of artificial intelligence (AI) and machine learning (ML) are becoming increasingly important in science. The successful application of these methods requires very large amounts of computing power. Since normal CPUs usually cannot deliver this level of performance, many operators have turned towards specialized accelerator chips for higher speeds and energy efficiency.
As the first scientific computing center in Germany, NHR@KIT has added a Graphcore IPU-POD16 system to the growing list of systems in its so-called Future Technologies Partition hardware and software testbed. The Future Technologies Partition enables researchers to port, test and run their scientific applications on a wide range of the latest, innovative and disruptive hardware architectures and technologies that are not part of the larger production systems yet.
Graphcore’s IPUs (Intelligence Processing Unit) feature a Multiple-Instruction Multiple-Data (MIMD) design, enabling fine-grained parallelism over multiple dimensions when training AI models. This new chip architecture lends itself to a wide range of new and emerging computational techniques in AI, such as the use of sparse data structures – common in artificial intelligence – where information cannot be processed in dense, contiguous blocks. The IPU’s broader range of AI-centric capabilities is enabling AI researchers to explore novel approaches to problem solving, not possible using other technologies.
“Karlsruhe Institute of Technology has always been at the forefront of innovation, providing our users with the tools and technologies that unlock new computational possibilities,” said Simon Raffeiner, Deputy Leader of the Core HPC Facility at Karlsruhe Institute of Technology.
“Graphcore systems, powered by the Intelligence Processing Unit, have been designed specifically for AI/ML and to enable a wider range of applications and approaches than are currently possible. Rather than limiting the ambitions of our researchers, IPU technology encourages them to explore and drive the field forward.”
The Graphcore IPU-POD16 system installed at NHR@KIT delivers 4 PetaFLOPS (FP16) AI compute power with its 16 IPUs and features 1038.4 GB of total memory, with 14.4 GB of that being in-processor memory. Graphcore’s Poplar SDK supports all major AI frameworks such as TensorFlow, PyTorch, PyTorch Lightning, Hugging Face, Keras, ONNX or PaddlePaddle, so existing AI models can be easily ported from other compute platforms to IPUs.
Researchers from all across Germany can submit their project proposals for the resources at NHR@KIT and be granted direct and dedicated access to the system. Two different types of project proposals can be submitted: Test projects can be used for the purpose of testing, porting, developing or benchmarking applications and are granted following a light-weight review process. Full projects in the range of tens of millions of CPU hours are reviewed by the scientific steering committee. While project proposals are being reviewed, preliminary access can be granted for setup and testing purposes.
Submitted by Thorsten Papkalla