Since coming online in the fall of 2019 in Paris, the Jean Zay supercomputer, named after the French politician because of his contributions to the scientific and arts fields, has been one of Europe’s most powerful supercomputers.
Primarily used in artificial intelligence (AI) and high-performance computing (HPC) research, the supercomputer has recently been beefed up with the addition of new Nvidia A100 80 GB GPUs.
Now, it has just become one of the first HPC machines to be integrated with a photonic co-processor, which uses light to transmit and process information.
Optical advantage.
Unlike traditional processors which use electric current, Optical Processing Units (OPU) by LightOn, a company specialising in photonics processing, transmit data using light.
At the most basic level, OPUs encode data as the brightness of light, which is manipulated using modulators. An OPU uses laser light that is shone onto a digital micromirror device to encode for 1s and 0s. The light is then redirected through a lens and random scattering medium assembly before being polarised and read by a conventional camera. This allows very large matrices to be manipulated in parallel.
OPUs integrated into the Jean Zay supercomputer use photonics to accelerate randomised algorithms at a very large scale, while working in tandem with standard silicon CPU and Nvidia’s A100 GPU tech – achieving “optical advantage”. With the explosion of data processing requirements today – fuelled by recent advancements in AI and machine learning applications – modern HPC is becoming more heterogeneous. As a result, this leads to HPC ecosystems featuring an array of specialised architectures and accelerators, such as CPUs, GPUs, TPUs, DPUs, OPUs, all working together to deliver optimal performance in data-intensive research applications.
One research area that particularly benefits from the integration of photonics processing is computational chemistry. For instance, Molecular Dynamics studies that underpin research in SARS-CoV-2 were noticeably enhanced by OPUs due to their extremely low latency times.
Pushing the frontiers of HPC.
LightOn’s Appliance integrated computing unit, powered by their Aurora2 OPU, is built into a 2U form factor to enable quick and easy integration into supercomputers. It can reach a peak performance of 1.5 PetaOPS at 30 W TDP. LightOn also claimed that its Appliance co-processor can deliver performance that is 8 – 40 times higher than GPU-only acceleration. Furthermore, with no electrical current used in data processing, this means there is no Ohmic heating for photonics co-processors. This could also potentially translate to higher efficiency and improved operating costs.
The new tech will be offered to some users of the Jean Zay supercomputer, who will carry out research on machine learning foundations, differential privacy, satellite imaging analysis and natural language processing tasks.
Together with the nascent quantum computing industry, the integration of photonics-based processors could usher in the new era of hybrid computing, offering highly performant systems with better energy efficiency.
