EuroEXA > Ambition

Setting out our ambitions 

To achieve our vision of a platform that demonstrates the potential for ExaScale performance with next-generation energy efficiency, we set out a range of ambitious key project deliverables:

ExaScale Deployment Recommendations

EuroEXA worked to develop an innovative system architecture, and show how it can be scaled to production ExaScale performance for real-world applications.

As such, set out to build, demonstrate and analyse physical testbed systems, and use these testbeds to explore how our architecture performs when scaled to ExaScale. We did this using simulation and modelling, based on a range of performance metrics and instrumentation from our applications on the EuroEXA prototypes and state-of-the-art HPC systems.

Benchmark Results

The high-performance conjugate gradient (HPCG) is not only a vital benchmark for the TOP500 List but also demands computation patterns close to real-world applications – as opposed to the high-performance LINPACK (HPL) benchmark.

The HPCG benchmark focuses on measuring the capability of a supercomputer to solve a large, sparse linear system. These numerical operations are usually found in scientific computing applications that are memory bound, often featuring low arithmetic intensities, irregular data access patterns and neighbour-collective communications – which, therefore, makes them extremely difficult to optimise for modern supercomputing architectures. 

We have developed an OpenCL version of the HPCG benchmark – based on HPCG CPU v.3.1 – and worked to optimise this on our own reprogrammable compute accelerators, also referred to as FPGA.

Testbeds

Over the life of EuroEXA, we built a series of testbeds to help us demonstrate the effectiveness and performance of our technologies, as the project developed. Primarily installed at the STFC Daresbury Laboratories in the UK, our testbeds offer increasing complexity with each iteration, allowing us to benchmark performance, demonstrate scalability and test every aspect of the system against real-world use-cases.

Public Deliverables

Confidential Deliverables

These are deliverables the project will be producing, but will not be made public due to reasons around commercial and intellectual property.