Supercomputing wins are a big deal for semiconductor manufacturers. Winning space in top systems is seen as a mark of stability and longevity. It implies vendors at the top of the market and with access to important, long-term government contracts believe your hardware is robust and capable enough to be used for cutting-edge scientific research at one of the United States’ largest installations. The economics of these deals are decidedly more cloudy, but as a matter of public relations, companies tend to adore them. Both Nvidia and AMD have something to crow about today, given a new announcement from Cray and Berkeley National Labs.
The National Energy Research Scientific Computing Center will use a Cray Shasta installation for its next-generation supercomputer, codenamed “Perlmutter.” To be clear, “Shasta” is what Cray calls the supercomputer architecture, while Perlmutter is the name of the specific NERSC system to be constructed. Shasta is designed to be compatible with both ARM and x86 architectures from Intel and AMD with support for a variety of interconnect standards, including Cray’s Slingshot, Intel’s Omnipath, and Mellanox (Infiniband). For those of you who associate Infiniband with Intel, the company bought its own Infiniband technology back in 2012, but Omnipath is a different, Intel-specific technology, which Mellanox’s Infiniband competes with. Shasta is designed to integrate components and accelerators from a variety of companies, including GPUs, FPGAs, and AI-specific accelerators.
Cray’s new Slingshot interconnect is a critical part of the system (and scaling interconnect power downwards is important to how we eventually reach exascale compute capability).