• CoreWeave and Parallel Works have deployed a managed AI computing environment for DARPA's NODES research programme using NVIDIA H100 systems
  • Each research team gets guaranteed computing capacity and can use additional shared resources when they are available

The fact

CoreWeave and Parallel Works have deployed a managed AI and high-performance computing environment for DARPA's Network of Optimal Dynamic Energy Signatures, or NODES, programme. Researchers access the system through Parallel Works' ACTIVATE platform, which handles sign-on, user accounts, provisioning, scheduling and usage reporting. CoreWeave provides the underlying computing, storage and networking, including NVIDIA HGX H100 systems and high-speed InfiniBand connections.

The capacity is divided between the NODES research teams. Each team receives a guaranteed allocation and can also use shared capacity when it is available. The companies also provide technical support, so individual teams do not have to build and manage separate cloud environments themselves.

NODES is developing deep-learning tools to analyse protein sequences and predict biological functions. DARPA allows researchers to use their own computing resources, but says final products and capability demonstrations must run on government high-performance computing systems. The new cloud environment supports the research programme; it does not show that NODES has achieved its scientific goals.

The assessment

Research teams do not just need GPUs; they need a practical way to get access to them. In this case, Parallel Works handles accounts, access and scheduling, while CoreWeave supplies the computing platform. That should leave NODES researchers with less infrastructure to set up and manage before they can start their work.

The way the capacity is shared will matter. Each team has a guaranteed allocation, with extra capacity available from a common pool when it is free. That gives researchers something they can plan around while making better use of spare resources. The question is how well that arrangement holds up when several teams need more computing power at the same time.

For BTW readers, this is a useful example of AI infrastructure being delivered as a managed service rather than simply rented as hardware. Its value will become clearer if researchers can get onto the system quickly, run their work without long waits and spend less time dealing with the computing environment itself.

What to watch

Watch how quickly research teams are brought onto the platform and how the shared capacity performs as demand rises. Reporting on waiting times, resource use, failed jobs or technical support would show whether NODES teams are getting predictable access in practice.