Intel’s AI ASIC strategy will be based on Habana chips from now on.
In a move widely speculated to have been looming, Intel has axed Nervana’s NNP-T and NNP-I training and inference chips for the data center in favor of Gaudi and Goya chips from recent acquisition Habana Labs.
A statement emailed to EETimes said that Intel will cease development on Nervana’s NNP-T AI training chip (Spring Crest) for the data center, while merely honoring existing customer commitments to the NNP-I inference chip (Spring Hill), following “customer feedback”.
“After acquiring Habana Labs in December and with input from our customers, we are making strategic updates to the data center AI acceleration roadmap. We will leverage our combined AI talent and technology to build leadership AI products,” Intel’s statement said. “We will bolster the current and next generation of Habana Goya and Gaudi with Intel’s AI hardware and software innovations. The Habana product line offers the strong, strategic advantage of a unified, highly programmable architecture for both inference and training. By moving to a single hardware architecture and software stack for data center AI acceleration, our engineering teams can join forces and focus on delivering more innovation, faster to our customers.”
Habana Labs, which was acquired by Intel last month for close to $2 billion, produced impressive results in the latest round of MLPerf inference benchmark scores. Habana’s results were second only to Nvidia’s in some categories and it seems Intel’s customers took notice, possibly even voicing their feelings to the chip giant on whose architecture they really wanted.
A key ingredient in Habana’s secret sauce is its on-chip RoCE (Remote direct memory access over Converged Ethernet) network. Competing solutions need an extra chip for this; Nvidia is in the process of acquiring Mellanox for $6.8bn for exactly this technology. Having RoCE on-chip makes for a level of scalability that suits the rapidly increasing size of neural network models.
In contrast to its strategy with Nervana, which was fully ingested by Intel, Habana will remain an independent business unit led by the same management team.
Intel also said in its statement that there would be no change to its roadmap for Movidius, which makes ultra-low power AI accelerators for computer vision in applications such as drones. Intel also offers CPU accelerators with its Xeon line, plus FPGAs from Altera, both targeted at the data center. The company is also working on a general-purpose GPU (Xe) which is anticipated later in 2020.
Founded in 2014, Nervana, based in San Diego, was purchased by Intel back in August of 2016 for a sum believed to be around $400 million. It was co-founded and led by Naveen Rao, who remains Intel’s corporate vice president and general manager of the AI Platforms Group.
Nervana’s NNP-I and NNP-T were launched with great fanfare in November 2019. The inference chip, NNP-I offered up to 50 TOPS for INT8 calculations. The training device, NNP-T promised performance up to 119 TOPS with 24 Tensor processing cores and 60 MB on-chip SRAM and 32GB high-bandwidth memory (HBM). It also featured an interconnect fabric called inter-chip link (ICL) which connected chips on the same or different cards without additional switching hardware, for the purposes of scaling.