Intel-owned Habana Labs' data center AI accelerators are already in the cloud, next they’ll appear in a brand new supercomputer.
Habana Labs’ AI training and inference accelerators have been selected to power AI compute in the Voyager supercomputer at the San Diego Supercomputer Center (SDSC) at UC San Diego. The new system will use 336 Habana Gaudi AI training accelerators and 16 Goya AI inference accelerators.
Habana announced a year ago that its data center chips are already used in the cloud at AWS,but this is the first public reveal of an HPC application for the Intel-owned company.
Voyager will be optimized for deep learning and AI workloads so it can be dedicated to AI research across many science and engineering fields. Habana’s Gaudi training processors are built for AI compute at scale with a solid combination of processing and networking capabilities (ten on-chip 100-Gigabit Ethernet RoCE v2 ports).
Voyager’s Habana-powered systems will be supplied by server maker Supermicro. The system will include Supermicro Gaudi AI training system boards, each with eight Habana Gaudi training chips paired with dual-socket 3rd generation Intel Xeon Scalable processors (Ice Lake), which were announced earlier this week. There will also be Supermicro SuperServer Goya PCIe cards for AI inference acceleration, which each have dual-socket 2nd gen Intel Xeon Scalable CPUs.
Voyager, which is funded by a $5 million grant from the National Science Foundation, is scheduled to be up and running in by autumn 2021. Its first three years of operation will be a testbed phase, during which SDSC will work with select research teams from astronomy, climate sciences, chemistry, particle physics and other fields to gain AI experience and insights leveraging Voyager’s unique features. SDSC will expand Voyager’s user base in years four and five.
This article was originally published on EE Times.