Intel launches Gaudi2 processor for AI deep learning

Intel launched its second-generation Habana Gaudi2 AI deep learning processor at the start of its Intel Vision 2022 event on Tuesday to help customers with AI training in data centers.

The chipmaker also introduced the Greco processor for inference work, a successor to the Goya processor. It runs on 75 watts of power, down from 200 watts in Goya. It will be sampled to select customers in the second half of 2022.

In addition, Intel launched new 12th Gen Intel Core HX processors for hybrid workstations, providing up to 16 cores and speeds up to 5 GHz.

Gaudi2 processing chops

Gaudi2 showed nearly double the performance on computer vision and natural processor throughput of Nvidia’s A100 80 GB processor, Intel said. The company also claimed up to 40% better price performance over existing GPU systems in the AWS cloud.  Its maximum power consumption is 600 watts, up from 350 watts in the first version.

Habana Labs, an Intel unit, launched the first-generation Gaudi AI training processor in June 2019.  Gaudi2 relies on 7 nm process technology, down from the first gen 16 nm, and more than doubles the Tensor processor cores to 24.  In package memory is tripled to 96 GB and on-board SRAM is doubled to 48 MB.

First-gen Gaudi customers include Mobileye and Leidos who said it has lowered costs. “We expect Gaudi2, building on the speed and cost-efficiency of Gaudi1, to provide customers with dramatically accelerated model training, while preserving the deep learning efficiency we experienced with first-gen Gaudi,” said Cheton Paul, the CTO of health and services at Leidos.

gaudi2 processor in mezzaine card
Intel Mezzanine Card with Gaudi2 processor (Intel)

Intel IPU roadmap

While Gaudi2 may have been the biggest news out of the Vision opening day, the company also unveiled its roadmap through 2026 for its infrastructure processing units.

 IPUs are programmable networking devices for use by enterprises, cloud and communication service providers to free up performance for CPUs. “With an IPU, customers can better use resources with a secure, programmable, stable solution that enabled them to balance processing and storage,” said Patty Kummrow, general manager of the network and edge group at Intel, in comments to reporters.

For 2022, Intel’s second generation IPUs include Mount Evans (an ASIC IPU) and Oak Springs Canyon (a second gen FPGA IPU) will ship to Google and other service providers.

For 2023-2024, Intel’s third generation IPUs will be 400 GB, code-named Mount Morgan and Hot Springs Canyon, which are expected to ship to customers.

For 2025-26, Intel expects its next generation will be 800 GB IPUs expected to ship to customers.

RELATED: Intel revenues drop 7% on PC weakness