Voyager, a supercomputer based on Intel’s Habana Gaudi processor, was launched this week at the University of California, San Diego.
Habana Labs, an Intel subsidiary that develops deep learning accelerators, announced last April that its artificial intelligence (AI) training and inference accelerator had been selected by the UC San Diego Supercomputer Center for its Voyager supercomputer. . It entered the testing phase recently after half a year later than the original scheduled date.
Image credit / Habana Labs
Voyager is a supercomputer based at UC San Diego with funding from the National Science Foundation (NSF). Voyager will be used for AI model development, training and testing in the fields of high energy physics, biology, genetics, materials science, atmosphere and astronomy.
In terms of specifications, it consists of a 42-node Supermicro X12 Gaudi AI training system, which contains 336 Intel Gaudi acceleration chips to increase the scale of AI training processing, and is equipped with 16 Goya acceleration chips to enhance the processing performance of the AI inference model. Each Gaudi processor has 10 built-in 100Gigabit Ethernet ports supporting RoCE v2, and the 42-seat training system is connected in series to the Arista switch by four 400Gpbs network connections.
The UC San Diego researchers also said that Voyager is currently running applications through Jupyter Notebook on Gaudi, as well as the TensorFlow and PyTorch frameworks.
Intel also announced Habana Gaudi2, an AI training processor, and Habana Greco, an AI inference processor, at last week’s Intel Vision 2022 conference. Habana Gaudi 2 is a 7nm process, provides a 24-core Tensor processor core, and has a built-in 96GB HBM2E high-bandwidth memory with a bandwidth of up to 2.45TB/s. Greco is Goya’s successor, offering LPDDR5 memory and 128 MB on-chip memory.