Give US open science more than ten times the computing of today's Perlmutter machine for simulation and AI, and analyse data streamed live from telescopes, fusion machines and sequencers.
Test system Cech delivered Jan 2026; full Doudna system due to be installed late 2026.
Updated 17 Mar 2026Checked 11 Oct0 updates this week
Vera CPUs and Rubin GPUs share memory over NVLink, so simulation codes and AI models move data between them without slow copies.
Dell's direct liquid cooling carries heat off the chips in water loops; air cannot remove heat fast enough from racks this dense.
The ESnet research network streams data from DOE telescopes, fusion devices and sequencers in, so results come back during the experiment, not weeks later.
Cech, a small liquid-cooled system on the previous chip generation, lets staff rehearse assembly, cooling and software before the full machine lands.
| Spec | Doudna (NERSC-10) |
|---|---|
| Performance vs Perlmutter | More than 10xR (reported) |
| Power vs Perlmutter | 2-3xR (reported) |
| Researchers served | 11,000+R (reported) |
| Processors | NVIDIA Vera Rubin CPU-GPUR (reported) |
| Network | NVIDIA Quantum-X800 InfiniBandR (reported) |
| Test system Cech, FP64 | 5.76 petaflop/sR (reported) |
| Test system Cech, GPUs | 144 Blackwell GPUs, 72 Grace CPUsR (reported) |
R reported by the company
US Department of Energy
Federal research, national laboratories, AI for science
Federal agency overseeing energy policy, nuclear research, computing and science. Operates 16 national labs; manages Genesis Mission ($5B+ AI-enabled science initiative).
