The Defense Advanced Research Projects Agency (DARPA) is shifting its computational focus from infrastructure management to biological modeling through a new partnership between Parallel Works and CoreWeave. By deploying a managed AI and high-performance computing (HPC) platform for the Network of Optimal Dynamic Energy Signatures (NODES) program, the companies aim to bypass the traditional months-long lead times required to build and scale AI infrastructure. This deployment integrates the Parallel Works ACTIVATE control plane with CoreWeave’s AI-native cloud to provide researchers with immediate, large-scale computing resources. The strategic move addresses a critical bottleneck in scientific research: the technical overhead of managing complex compute environments. For enterprise IT leaders, this deployment illustrates a growing trend toward specialized, fully managed AI clouds designed to support high-stakes, resource-intensive research workloads.
Parallel Works ACTIVATE and CoreWeave Infrastructure Integration
The deployment for the DARPA NODES program centers on a managed environment designed to accelerate the training of AI models and the execution of complex biological simulations. Parallel Works is utilizing its ACTIVATE platform to serve as the primary control plane, managing critical operational tasks such as user onboarding, identity and access management, provisioning, scheduling, and reporting. This layer is intended to provide researchers with end-to-end support, including around-the-clock access to experts for both the platform and specific researcher applications. By abstracting the underlying complexity, the companies claim researchers can begin workloads within days rather than the weeks or months typically required for infrastructure setup.
The physical compute layer is built on CoreWeave’s AI cloud, which provides the heavy-duty hardware necessary for advanced biological modeling. Specifically, the environment includes a dedicated reservation of NVIDIA HGX H100 systems, supported by enterprise-scale storage and high-speed NVIDIA Quantum InfiniBand networking. A key technical feature of this integration is the ability of the ACTIVATE platform to divide this dedicated reservation across various NODES research teams. This ensures each team maintains a guaranteed allocation of resources while retaining the ability to draw on shared capacity when it becomes available, optimizing the utilization of the high-cost GPU hardware.
Orchestration and Observability for the NODES Program
To manage diverse scientific workloads, the platform employs a unified orchestration strategy that allows researchers to access both SUNK and the CoreWeave Kubernetes Service (CKS) through a single sign-on. SUNK, which is CoreWeave’s Slurm on Kubernetes offering, is positioned to enable researchers to run workloads of varying sizes and types within the same underlying compute environment. Parallel Works ACTIVATE manages the specific accounts, allocations, and sharing policies that dictate how these various research teams interact with the shared hardware. This unified approach is intended to reduce the friction of switching between different orchestration tools during intensive research cycles.
Operational reliability is addressed through a dual-layered observability strategy. Parallel Works ACTIVATE provides the NODES program and individual research teams with detailed usage and utilization reporting to track resource consumption. Simultaneously, CoreWeave’s observability stack monitors the health of the underlying hardware fleet. The goal of this integrated monitoring is to identify and resolve infrastructure-layer issues before they can interrupt or jeopardize multi-week computational runs. By combining these visibility tools, the platform seeks to provide a level of stability required for the dependable, large-scale computing that the NODES program demands for studying the dynamics of biological systems.
Key Takeaways
- The NODES program will utilize a managed environment featuring NVIDIA HGX H100 systems, enterprise-scale storage, and NVIDIA Quantum InfiniBand networking.
- Parallel Works’ ACTIVATE platform manages user onboarding, identity, provisioning, and scheduling, while providing 24/7 direct-to-expert support.
- Researchers can access both CoreWeave Kubernetes Service (CKS) and SUNK (Slurm on Kubernetes) through a single sign-on via the ACTIVATE control plane.
TechInsyte's Take
In our view, the collaboration between Parallel Works and CoreWeave signals a significant shift in how high-stakes government and scientific research will consume AI resources. Rather than attempting to build bespoke, in-house supercomputing clusters—a process fraught with procurement delays and technical debt—agencies like DARPA are increasingly looking toward "AI-native" cloud providers that can deliver specialized hardware like NVIDIA HGX H100s with immediate availability. This move highlights the growing importance of the "control plane" in the enterprise AI stack; having the raw GPU power is no longer enough if the orchestration, identity management, and observability layers are not seamlessly integrated. For the broader enterprise, this suggests that the future of specialized AI workloads lies in managed, highly orchestrated environments that prioritize researcher uptime over infrastructure ownership.
Questions & Answers
How does the ACTIVATE platform manage resource distribution among different research teams?
The ACTIVATE platform divides the dedicated reservation of NVIDIA HGX H100 systems across the NODES research teams. This configuration ensures that each team holds a guaranteed allocation of compute resources while also allowing them to access shared capacity when it is available.
What specific hardware and networking technologies are included in this deployment?
The managed environment includes a dedicated reservation on NVIDIA HGX H100 systems. This compute capacity is supported by enterprise-scale storage and high-speed NVIDIA Quantum InfiniBand networking to facilitate large-scale biological modeling and AI training.
How does the platform handle different types of computational workloads?
The platform uses unified orchestration through the ACTIVATE control plane, allowing researchers to access both CoreWeave Kubernetes Service (CKS) and SUNK (CoreWeave’s Slurm on Kubernetes offering). This enables the efficient execution of various workload sizes and types within a single underlying compute environment.
What mechanisms are in place to ensure the stability of long-running simulations?
The deployment utilizes a two-part observability strategy: Parallel Works ACTIVATE provides usage and utilization reporting for teams, while CoreWeave’s observability stack monitors the health of the underlying hardware fleet to catch infrastructure issues before they impact multi-week runs.
Source: Businesswire
