Nvidia wins first DOE award since 2022, with Doudna supercomputer (19459000)
US Department of Energy’s next supercomputer is being built by Dell Technologies, and will be powered by Nvidia’s next-gen Vera-Rubin accelerations. This is a significant change from the usual Cray-AMD tag team that builds such machines. This is Nvidia’s first DOE win since the Venado system of 2022. The Doudna System, named after Jennifer Doudna (Nobel laureate and pioneer of CRISPR gene-editing technology), is set to debut at the Lawrence Berkeley National Laboratory, California, next year.
The system promises a 10x improvement in “scientific output” over its predecessor, Perlmutter. It will also consume 2-3x less power.
At first glance, this suggests that the system could squeeze 790 petaFLOPS in double precision performance out of between 5.8 to 8.7 megawatts, making it by a large margin the most energy-efficient supercomputer ever. This is unlikely to be the case. While we don’t yet know how many Doudna systems will be equipped with the GPU giant’s next generation Vera-Rubin chips, or even the number of them, we do know that Nvidia’s Blackwell Ultra accelerators have sacrificed double precision performance, which was long considered crucial for scientific computing. Instead, they are using 4-bit precision formats designed to handle AI workloads.
Nvidia GB300 NVL72, a Blackwell Ultra-based system with 72 GPUs, can only produce 100 teraFLOPS in FP64 vector performance. A single AMD MI300A in Lawrence Livermore National Labs’ El Capitan system can achieve 61.3 double precision TeraFLOPS.
While we can’t guarantee that Nvidia’s Rubin Accelerators will continue this trend in the future, we suspect this may be why Nvidia claims a 10x increase on performance rather than “scientific output” .
Having said that, the Doudna System may not require a lot of 64-bit FLOPS in order to complete its mission. The machine is designed to be a Swiss Army Knife capable of running a wide range of workloads, including traditional HPC and AI. Speed is also a priority. NERSC Director Sudip dosanjh made 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.”
Researchers plan to use data streaming from the DIII-D National Fusion Ignition Facility in San Diego to perform real-time modeling of plasma.
The system will be equipped with Nvidia’s Quantum X InfiniBand network, which will provide up to 800 Gb/s per port – 4x faster than Slingshot NICs currently used in DoE supercomputers. Nick Wright, Doudna’s chief architect, added. “Now it’s part of the entire workflow connected to experiments, telescopes, and detectors.”
It is expected that the system will serve approximately 11,000 researchers, including those working in fusion power, materials science, drug discovery, astronomy, and protein design. Senators are upset that Nvidia has plans to build a Shanghai R&D laboratory. They also claim that Dell has a $14BN AI server backlog. The National Energy Research Scientific Computing Center has used AI in the past to predict novel proteins structures, analyze proton-data from particle accelerators, or model complex chemical reactions.
This is where Nvidia’s Vera Rubin accelerators are expected to shine. We learned at GTC in the spring that each dual GPU package would deliver 50 petaFLOPS FP4 computation and feature 288GB of fast HBM4 memory.
Doudna, along with AI, will also support research on quantum computing algorithms via Nvidia’s CUDA Q. The dev platform allows users to simulate workloads on conventional CPU or GPU hardware, or deploy them on quantum processors. (r)
