Customer spotlight
Customer: National Laboratories
Cerebras collaborated with researchers at Sandia National Laboratories, Lawrence Livermore, Los Alamos National Laboratory, and the National Nuclear Security Administration on this record setting result to unlock the millisecond-scale for scientists, enabling them to see further into the future than ever before.
Use Case
Computational fluid dynamics
Computational fluid dynamics (CFD) codes are a central component of aerospace and energy research. These workloads require massive sparse computation at extraordinary memory and communication bandwidth. Researchers typically take great lengths to develop state of the art implementations on massive supercomputers. The Cerebras WSE-3 puts all of these resources on a single chip, delivering supercomputer-impossible performance of 100s-10,000s of legacy C/GPU machines for this work.
Working with researchers at the National Energy Technology Laboratory (NETL), Cerebras showed that a single Cerebras system can outperform one of the fastest supercomputers in the US by more than 200X on a stencil computing workload.
use case
AI-accelerated modeling & simulation
Traditional HPC can be accelerated by AI surrogate models that learn physics or by AI models that augment traditional supercomputers to inform the next steps of a physics-based simulation code. As the world’s fastest AI computer, the CS-3 is uniquely well-suited as a specialist accelerator for this type of work in a heterogenous AI+HPC cluster.
In addition, the Cerebras architecture can massively accelerate sparse linear algebra and tensor workloads, stencil-based partial differential equation (PDE) solvers, N-body problems, and signal processing algorithms such as FFT.
Use Case
Molecular dynamics
Molecular dynamics simulations power a range of applications from drug discovery to materials science. Calculating the time dependent state and interactions across large-scale simulations typically requires hundreds of nodes on a supercomputer. Deep neural networks are increasingly used for the prediction of energies and forces in molecules. With the CS-3, researchers can train physics-informed neural networks and accelerate MD simulations codes on a single machine with the WSE-3’s high bandwidth local on-wafer SRAM and interconnect.
Testimonial
"The CS-1 allowed us to reduce the experiment turnaround time on our cancer prediction models by 300X, ultimately enabling us to explore questions that previously would have taken years, in mere months."
