King’s College London

Phase One - NVIDIA DGX Infrastructure

On the 23rd July 2019, Scan AI, NVIDIA, Mellanox and NetApp came together at Kings College London to reveal the collaborative deep learning & AI solution that had been deployed in KCL's Medical Imaging & AI Centre. The projects aim was to bring artificial intelligence in medical imaging to the point of care. In contrast to traditional medical testing, which involves sending scans for further analysis by specialists, point of care testing allows the results gained from X-rays, CT scans or MRI to be delivered immediately at the time of the patient-doctor interaction.

The collaborative partners deployed the NVIDIA DGX-2 AI supercomputer as part a cluster that powers the first phase of the project. The system’s large memory and two petaflops of compute prowess enables the KCL team to tackle the training of large, 3D datasets in minutes instead of days. The work aims to discover breakthroughs in classifying stroke and neurological impairments, determining the underlying causes of cancers, as well as recommending the best treatment for patients.

Phase Two - Run:AI Implementation

To further increase GPU utilisation within the NVIDIA DGX-1 and DGX-2 appliances, the AI Centre recently installed Run:AI, the world’s first orchestration and virtualisation platform for AI infrastructure. Run:AI pools computing resources to optimise GPU utilisation and elastically allocates hardware to the tasks that need it. It allows data scientists to use massive amounts of GPU power to accelerate their research while giving IT teams control and real-time visibility over resource provisioning, queuing and utilisation.

The Run:AI platform has helped the AI Centre to run important research into the progression of neurological conditions and image-guided interventions for cancer patients. It has also allowed the team to quickly pivot research activities towards COVID-19, where time to results is critical.

Phase Three - Introduction of Federated Learning

Patient data, by its nature, is required to be kept anonymised at every stage of the deep learning process, so the challenge is to ensure maximum learning is gained from the data but without compromising its integrity. Federated Learning is a machine learning technique that trains an algorithm across multiple decentralised edge devices or servers holding local data samples, without exchanging them. In practice this means a centralised AI algorithm can learn from locally stored anonymised data at multiple NHS Trusts by building a consensus model, from multiple data sources, without any patient data leaving the secure data enclaves at each hospital.

The AI Centre at KCL is now in partnership with eleven other NHS Trusts, and in the process of deploying an ambitious Federated Learning Interoperability Platform (FLIP), to support pioneering artificial intelligence systems across the NHS. This will be supported by the role out of AI-optimised POD architectures at multiple sites.

Phase Four - DGX POD Installation

In order to establish federated learning model across multiple sites, there is a need for the hardware capability to be standardised across all the participating sites so no performance bottlenecks are generated in the system. This has been achieved by the role out of ‘POD’ infrastructures in each location. These pods consist of GPU-accelerated server hardware, storage hardware to handle local data and high throughput networking to keep data rapidly available to the compute nodes.

A combination of NVIDIA DGX and 3XS EGX servers were used to get the required compute performance whilst remaining cost-effective enough that multiple units can be installed within a short time frame for maximum efficacy of the federated learning platform.

Phase Five - Update of KCL Cluster

Following the successful launch of DGX PODs across the multiple federated learning sites, it was decided that the master KCL DGX cluster should be relocated to a London datacentre for improved management and monitoring. The Scan AI team undertook a project to transport and install the high availability DGX cluster hardware into new racking, configure the NetApp A800 storage system and update the NVIDIA network switches.

Once the hardware was in position and tested, further work was undertaken to reconfigure the Kubernetes software and deploy a new install of the Run:ai Atlas GPU virtualisation platform.