GigaIO Reveals Latest Breakthrough in Single-Node GPU Power
Note: This article originally appeared on hpcwire.com, click here to read the original piece.
November 10, 2023
SAN DIEGO, Nov. 10, 2023 – GigaIO, a leading provider of workload-defined infrastructure for AI and accelerated computing, has announced the launch of the world’s first 64 GPU single-node supercomputer, called SuperDuperNODE. GigaIO introduced a 32-GPU SuperNODE earlier this year, at the time breaking records with 32 GPUs on a single node. With the launch of the SuperDuperNODE, customers working in generative AI and HPC can now look forward to doubling performance.
In the field of HPC, where more than half of all code is estimated to run on a single node, SuperNODE and SuperDuperNODE are making possible what was previously unthinkable — for example, running the world’s largest Computational Fluid Dynamics (CFD) simulation with FluidX3D in just 33 hours.
That simulation, run by Dr. Moritz Lehmann, the developer of FluidX3D ProjectPhys, shows the 62m-long Concorde aircraft before landing at 300km/h airspeed and 10° angle of attack, for one second in flight. The simulation resolution is an amazing 40 billion cells, with a tiny cell size of (12.4mm)³. 67,268 time steps were computed in 29 hours, plus four hours for rendering 5×600 4K frames, for a total runtime of 33 hours. The same simulation run on traditional servers would need several months of compute time, whereas with the original SuperNODE, it was run over a single weekend.
Several Top500 benchmarks for supercomputer performance testing also demonstrated extraordinary performance:
- HPL-MxP showed excellent scaling and reduced precision compute bandwidth running on the SuperNODE, achieving 99.7% of ideal theoretical scaling.
- HPL testing resulted in 95.2% of ideal theoretical scaling.
- HPCG showed 88% scaling, an excellent result for memory scaling.
These results demonstrate that the power of supercomputers many times the size, cost, and complexity of SuperNODE can be accessed in an easy-to-deploy, seamlessly scalable, and cost-effective engineered solution. In particular, the HPL-MxP result indicates SuperNODE is an excellent vehicle for machine learning and AI workloads.
“GigaIO is a valued partner in SourceCode’s co-design strategy to deliver next-generation AI-enabled infrastructure to the market,” said Eric Sindelar, EVP Business Development and Partner Alliances at SourceCode. “The insatiable appetite for GPUs and memory in AI applications requires powerful new solutions. With 64 GPUs in a single node, the SuperDuperNODE platform establishes a new standard for high-end AI and fills a crucial gap in the market.”
“I would like to praise the work of our technical team and that of our partners — AMD, Dell, and Supermicro — for working together to resolve a long string of issues that have prevented systems like this from existing until today,” said Alan Benjamin, CEO of GigaIO. “The entire software and hardware stack have finally come together, and the result is a mind-blowing capability of 64 GPUs within a single node. Our universal dynamic memory fabric, FabreX, enables users to seamlessly switch between using SuperNODE or SuperDuperNODE in Beast mode (one large server), in Swarm mode (many servers operating together), or in Freestyle mode (everyone gets their own server). The system finally conforms to the workload rather than the other way around — no matter what changes may come along at a later date.”
The GigaIO SuperNODE is shipping now, available directly from Dell, Supermicro, and selected channel partners, and the SuperDuperNODE will ship by the end of the year.
Learn more about the SuperNODE at the GigaIO booth #2031 and see it in the Arc Compute booth #633 at SC23 in Denver.
About GigaIO
GigaIO provides workload-defined infrastructure through its universal dynamic memory fabric, FabreX, which seamlessly composes rack-scale resources and integrates natively into industry-standard tools. The SuperNODE and the SuperDuperNODE are “impossible servers,” fully engineered to “Just Work” for AI and accelerated computing. These solutions allow users to deploy systems in hours instead of months and run more workloads at lower cost through higher utilization of resources and more agile deployment.
View source version on hpcwire.com: https://www.hpcwire.com/off-the-wire/gigaio-reveals-latest-breakthrough-in-single-node-gpu-power/