Nvidia scores its first DOE win since 2022 with Doudna supercomputer : US Pioneer Global VC DIFCHQ SFO NYC Singapore – Riyadh Swiss Our Mind

System promises a 10x increase in ‘scientific output’ – not necessarily performance

The US Department of Energy’s next supercomputer will be built by Dell Technologies and powered by Nvidia’s next-gen Vera-Rubin accelerators – a notable switch from the usual Cray-AMD tag teams that build such machines. It’s the first DOE win for Nvidia since the Venado system in 2022.

Named for Nobel laureate Jennifer Doudna, who pioneered CRISPR gene editing tech, the Doudna system is set to make its debut at the Lawrence Berkeley National Laboratory in California next year.

Compared to its predecessor, Perlmutter, the system promises to deliver a 10x increase in “scientific output” while consuming just 2-3x the power.

On first blush, that suggests the system should squeeze 790 petaFLOPS of double precision performance from between 5.8 and 8.7 megawatts of power, making it the most power-efficient supercomputer on record by a wide margin.

However, that probably won’t end up being the case.

While we still don’t know much about the GPU giant’s next-gen Vera-Rubin superchips or even how many the Doudna system will feature, we do know Nvidia’s latest batch of Blackwell Ultra accelerators sacrificed double-precision performance, long considered essential for scientific computing, in favor of increased use of 4-bit precision formats tailored to AI workloads.

Nvidia’s Blackwell Ultra-based GB300 NVL72 with 72 GPUs on board churns out just 100 teraFLOPS of FP64 vector perf. For reference, a single AMD MI300A found in Lawrence Livermore National Lab’s El Capitan system is good for 61.3 double precision teraFLOPS.

We can’t be sure Nvidia’s Rubin accelerators will continue this trend, but we suspect this is why Nvidia is claiming a greater than 10x increase in “scientific output” rather than performance.

With that said, the Doudna system may not need a ton of 64-bit FLOPS to complete its mission. That’s because the machine is envisioned as something of a Swiss Army knife capable of running a variety of workloads that span traditional HPC and AI, with speed being a key priority.

“The Doudna supercomputer is designed to accelerate a broad set of scientific workflows,” NERSC Director Sudip Dosanjh said in a statement. “Doudna will be connected to DoE experimental and observational facilities through the Energy Sciences Network (ESnet), allowing scientists to stream data seamlessly into the system from all parts of the country and to analyze it in near real time.”

For example, researchers plan to perform real-time plasma modeling using data streamed from the control room of the DIII-D national fusion ignition facility in San Diego.

To support this, the system will feature Nvidia’s Quantum-X InfiniBand networking, which will deliver up to 800 Gb/s of bandwidth per port — 4x faster than the Slingshot NICs used in DoE’s most powerful supers today.

“We used to think of the supercomputer as a passive participant in the corner,” Doudna chief architect Nick Wright added. “Now it’s part of the entire workflow connected to experiments, telescopes, and detectors.”

The system is expected to serve roughly 11,000 researchers in the areas of fusion power, materials sciences, drug discovery, astronomy and protein design to name just a few.

Much like El Capitan, applying AI to scientific workloads remains a major focus for the DoE. The National Energy Research Scientific Computing Center (NERSC) has previously used AI to predict novel protein structures, analyze proton data from particle accelerators, and model complex chemical reactions.

And this is where we expect Nvidia’s Vera-Rubin accelerators really shine. As we learned at GTC this spring, each dual GPU package will deliver 50 petaFLOPS of FP4 compute and will feature 288GB of speedy HBM4 memory.

Along with AI, Doudna will also support research into quantum computing algorithms through Nvidia’s CUDA-Q. The dev platform can either be used to deploy workloads on quantum processors, or simulate them on conventional CPU and GPU hardware. ®

https://www.theregister.com/AMP/2025/05/30/nvidia_doe_supercomputer/