Graphcore’s IPU has already made a huge impact on the AI world, accelerating workloads and opening up new avenues of exploration.
A growing number of users are also finding applications for our high performance, massively parallel processor beyond artificial intelligence – for example, in areas such as path (or ray) tracing, scattered data interpolation, and a number of scientific processes such as Kalman filters in particle physics.
We are delighted to be further supporting these efforts with the launch of the Accelerated Computing Academy—a programme to provide systems access and bespoke support for compute-intensive academic applications beyond machine learning.
The Accelerated Computing Academy is backed by a new European computing research alliance Graphcore has formed alongside top experts in academia and HPC including from the University of Bristol, Imperial College London, the University of Oxford, Simula Research Laboratory, and University College London.
The programme will offer a range of exclusive benefits including provision of free access to hardware, expert training workshops, dedicated engineering guidance, and broad support for the underlying research projects. Support will include letters for grant and funding proposals, and members will be featured in regular project showcases and developer spotlight promotions. Participants will also benefit from the support and guidance of leading academics with experience of developing novel applications for the Graphcore IPU.
Additionally, programme members will also be eligible to apply for dedicated internship opportunities (more details of which will be provided in due course).
“The Accelerated Computing Academy is a fantastic initiative that will doubtless yield many breakthrough applications of Graphcore’s Intelligence Processing Unit, just as those we are developing for the future of the Large Hadron Collider beauty experiment, where immense data rates are processed in real time to perform the most precise measurements on matter-antimatter asymmetries,” said Jonas Rademacker, Professor of Physics at the University of Bristol.
“I am looking forward to working with participants in the Accelerated Computing Academy; sharing insights and experiences of using the Graphcore IPU and helping to unlock a range of exciting new applications,” said Natalia Ares, Associate Professor in Engineering Science at the University of Oxford.
Together, we hope to help solve today’s most challenging problems and make a positive impact on the world by connecting computing visionaries with cutting edge hardware.
The Accelerated Computing Academy is open, from launch, to applicants worldwide.
Who is the Accelerated Computing Academy for?
The Accelerated Computing Academy is specifically designed for computer scientists in academia who are looking to solve new problems through computationally intensive research approaches outside of AI and machine learning. Of particular relevance is low-level programming using our native Poplar APIs, making direct use of all the IPU’s hardware capabilities.
This programme has been designed to complement other initiatives within our Academic Programme, which are primarily aimed at researchers using high-level machine learning frameworks such as PyTorch or TensorFlow.
We hope to work across the academic spectrum, with researchers in computer science, mathematical sciences, physics, data analytics, finance, energy, chemistry, aerodynamics, biology, geographical sciences, and more.
Use cases could include simulation modelling, structural analysis, video processing, computational finance, fluid dynamics, weather prediction, medical imaging, oil & gas exploration, renewable energy, financial security, and large-scale statistical analysis.
Aren’t Graphcore’s IPUs designed for AI and machine learning?
Although Graphcore’s IPU technology has been designed primarily for AI and machine learning workloads, it has tremendous potential to be used for scientific applications entailing highly parallel, high-performance compute, generating results faster while potentially consuming less energy.
Scientific software is designed from the ground up to be massively parallel, allowing it to scale to tens of thousands of nodes in a supercomputer. But complex data dependencies and memory access patterns in numerical algorithms mean that application performance is often limited by memory bandwidth or communication constraints rather than the speed at which individual nodes can compute. This is where Graphcore’s IPU offers exciting opportunities to unlock future performance.
Whereas CPUs are designed for scalar processes and GPUs are designed to handle large blocks of dense contiguous data, IPUs are massively parallel processors benefitting from a unique multiple instruction, multiple data (MIMD) distributed architecture with thousands of tiles (cores) per chip. Each tile in an IPU has its own memory, with large amounts of ultra-high bandwidth on-chip memory being placed adjacent to the many processor cores.
IPU tiles can exchange data extremely quickly under the high-performance Bulk Synchronous Communication (BSP) all-to-all communication architecture. With BSP, tiles alternate between local computation and data exchange with other tiles, with a synchronisation step in between. Graphcore’s Poplar SDK manages all communication between the different processor cores and between the different independent parallel programs.
This execution scheme enables IPUs to efficiently parallelise operations, even across hundreds of thousands of tiles. Such a level of explicit parallel execution control is missing from all other processors but is critical in making the execution of compute-intensive tasks robust and scalable across a large scale-out machine.
Co-designed alongside the IPU from the start, Graphcore’s Poplar SDK is a mature software stack that enables fine-grained programming at the lowest level to exploit tile parallelism and unlock the full performance of the hardware. Applications running on IPUs can be programmed easily, since Poplar provides a C++ interface for development and a complete set of debugging and analysis tools is available to help tune performance. Some particularly notable case studies of C++ application development are linked below.
All of this enables fine-grained parallelism, helping to achieve high throughput while using small convolutions and processing very few data samples in parallel.
How can I join?
You can apply today to join the Accelerated Computing Academy—find out more on the programme's dedicated webpage.
This programme is an extension of our Academic Programme initiatives to support research at the leading edge of scientific innovation. If you’re interested in advancing research in advanced machine learning and AI fields, please take a look at our Machine Intelligence Academy instead.
