This article has been updated from Dec. 3 to note Intel's confirmation of the Habana Labs deal.
Three years after it bought AI/deep learning chip startup Nervana Systems, Intel (INTC) is making an even costlier AI-related acquisition.
On Monday morning, the chip giant confirmed recent media reports by announcing that it has acquired Habana Labs, an Israeli chip startup that is six months removed from unveiling an AI training processor known as Gaudi.
Intel said it's spending about $2 billion to buy Habana. That figure is at the high end of a previously-reported range of $1 billion to $2 billion, and nearly 5 times the reported $408 million price attached to the Nervana acquisition. Habana is Intel's biggest acquisition since its $15.3 billion, early-2017 deal to buy top ADAS vision processor supplier Mobileye.
The acquisition is a little surprising, given that Intel has been busy working to commercialize a pair of Nervana server accelerators, and has also been making a lot of noise about these efforts.
For AI training -- the computationally-demanding job of creating deep learning models that can effectively do things such as translate text, decipher voice commands and analyze photos and videos -- Nervana has been prepping an ASIC that's now known as the NNP-T. For AI inference -- the running of trained AI models against new data and content, such as voice commands and photo/video uploads from consumers -- Nervana has been developing an ASIC known as the NNP-I.
Habana, meanwhile, unveiled an inference ASIC known as the Goya HL-1000 in September 2018, and followed that up by unveiling the Gaudi in June 2019. With the qualifier that rivals might beg to differ with these claims, Habana asserted in September 2018 that the Goya HL-1000 delivers "one to three orders of magnitude better performance than solutions commonly deployed in data centers today," and claimed in June that the Gaudi, which was then promised to sample in the second half of 2019, will "deliver an increase in throughput of up to four times over systems built with [an] equivalent number [of] GPUs."
Presumably, Habana was talking about Nvidia's (NVDA) server GPUs when making that claim. Nvidia has long been dominant in the AI training accelerator market, and its current flagship server GPU (the Tesla V100) is widely used by cloud giants and others to handle training workloads. In addition, while the inference market is more competitive, Nvidia's cheaper Tesla T4 GPU has begun seeing good traction in this space. Last month, the company reported that its inference revenue more than doubled annually during its October quarter.
Why did Intel want to buy Habana when it already has Nervana? One possibility is that Intel could be worried about how the NNP-T (codenamed Spring Crest) will stack up relative to Nvidia's next-gen flagship server GPU, and sees Habana's Gaudi delivering better price/performance. The GPU is generally expected to be based on a new architecture known as Ampere, rely on advanced 7-nanometer (7nm) manufacturing processes from Taiwan Semiconductor (TSM) and/or Samsung, and launch at some point next year.
In addition, startup Cerebras Systems recently made waves by unveiling its massive Wafer Scale Engine (WSE), which packs 400,000 processing cores and requires an entire 300mm chip wafer to produce. Last month, Cerebras revealed the CS-1, a hardware system that relies on the WSE, and disclosed that it has delivered a CS-1 system to the Department of Energy's Argonne National Laboratory.
In May 2018, Intel said Spring Crest, which like Habana's first chips is based on a relatively old 16nm TSMC manufacturing process, would arrive in late 2019. However, while Spring Crest has entered production and is sampling with cloud giants such as Facebook (FB) and Baidu (BIDU) , Intel now indicates a full commercial launch will happen at some point in 2020. For its part, server OEM Supermicro has said that it will offer systems featuring Spring Crest in mid-2020.
The innovative networking capabilities of Habana's Gaudi accelerator -- it comes with 10 100-gig Ethernet ports and supports a technology known as Remote Direct Memory Access (RDMA) -- might also be a selling point. Intel claims that AI training systems featuring large numbers of Gaudi nodes "are expected to deliver up to a 4x increase in throughput versus systems built with the equivalent number of GPUs."
Nvidia might beg to differ with that performance claim. And even if the claim held up, the numbers would presumably be different with Nvidia's next-gen, 7nm GPU.
Nonetheless, with Nvidia having inked a $6.9 billion deal to buy top high-speed server interconnect provider Mellanox Technologies (MLNX) (it's still awaiting Chinese and European approval), addressing server-to-server bottlenecks looks set to become a key battleground for training accelerator vendors, and Habana's technology helps Intel out here.
Throw in Nvidia's other competitive strengths, such as its large software ecosystem, long-standing relationships with AI R&D teams at cloud giants and extensive efforts to optimize the performance of its GPUs for various AI models and software frameworks, and you have a good set of reasons why Intel is again willing to turn to M&A to strengthen its position in a fast-growing AI accelerator market.