Berkeley Lab Takes Major Step Toward Doudna with Delivery of Early Access System, Cech

Key Takeaways

• Cech EAS installed at NERSC for Doudna preparation
• Features 72 Grace CPUs, 144 Blackwell GPUs, 5.76 PFLOPS
• Dell’s liquid‑cooling cuts cooling costs up to 60%
• Tests software stack (Omnia, OpenCHAMI) for mixed AI workloads
• Enables DOE Genesis Mission AI research before full system launch

Summary

Berkeley Lab’s NERSC has taken delivery of the Cech early‑access system, a scaled‑down prototype of the upcoming Doudna supercomputer. Built by Dell with NVIDIA Grace CPUs and Blackwell GPUs, the rack delivers 5.76 PFLOPS FP64 and 1.44 EFLOPS NVFP4 performance. The installation serves as a testbed for hardware, liquid‑cooling, networking and the complex software stack that will power Doudna later in 2026. NERSC will use Cech to validate processes, benchmark AI‑intensive workloads, and ensure a seamless user experience for its 11,000‑plus researchers.

Pulse Analysis

The Cech early‑access system marks a strategic milestone for the Department of Energy’s high‑performance computing roadmap. By delivering a compact version of the future Doudna platform, NERSC and Dell can fine‑tune delivery, rack integration, and power‑distribution workflows under real‑world conditions. The hardware mix—NVIDIA Grace CPUs paired with Blackwell GPUs—represents the cutting edge of AI‑ready compute, while the inclusion of NVIDIA Quantum‑X800 InfiniBand and VAST Data’s AI‑optimized storage creates a low‑latency, high‑throughput fabric that mirrors the full‑scale design. This hands‑on approach reduces the risk of costly re‑engineering once the full system arrives.

Beyond hardware, Cech serves as a sandbox for the sophisticated software ecosystem that will drive Doudna’s mixed workloads. Engineers are deploying the Omnia orchestration layer alongside OpenCHAMI to automate provisioning, telemetry, and QoS management across both traditional simulation and emerging AI pipelines. Early benchmarking on Cech will reveal performance bottlenecks, inform scheduling policies, and validate the VAST AI OS’s ability to balance scratch‑space speed with long‑term archival access. The system’s liquid‑cooling architecture, featuring Dell’s PowerCool eRDHx and Rittal’s in‑row coolant distribution, promises up to a 60% reduction in cooling energy, aligning with DOE’s sustainability targets.

For the broader scientific community, the successful validation of Cech accelerates the timeline for Doudna’s contribution to the DOE Genesis Mission and other AI‑intensive research programs. By proving that a high‑density, energy‑efficient platform can sustain exaflop‑scale AI workloads, NERSC positions itself as a model for other federal agencies seeking to modernize their HPC fleets. The lessons learned will inform future collaborations between national labs, hardware vendors, and software innovators, driving a more resilient and scalable HPC ecosystem across the United States.