Intel’s $2 billion purchase of Habana Labs for its AI capabilities will challenge Nvidia in the lucrative AI silicon marketplace, especially in the data center.
“Habana is really a shot across the bow against Nvidia more than anything, as Habana is a high performance, high end training solution accelerator that Intel claims is a 4x improvement over traditional GPU-based training systems,” said Jack Gold, an analyst at J. Gold Associates.
The investment is “key to Intel’s data center plans,” he added. Intel CPUs are “pretty good” at inference work, Gold said, “but not so good at teaching the AI models how to perform…Training is the heavy lifting in AI.”
“Training has been a hole in the Intel strategy,” Gold added. “In parallel with their FPGA’s that are also used for acceleration of specific AI functions, this is a good addition toward a more holistic strategy.” He called the Intel Xeon Scalable processors a “great AI engine for inference and simple training” but not sufficient for high end, large scale AI training which requires huge parallel processing.
That hole has been why Nividia has taken a lead in the market, Gold said, with its Tesla GPUs and the DGX workstation for the data center. It is why Google, Facebook, Amazon, Microsoft and others are building custom silicon.
So, with Xeon for inference and simpler training, Intel has several important pieces of the pie needed for a complete solution when adding in Intel’s high-speed memory, high performance networking, AI accelerators from Nervana and FGPAs from Alterra. What’s missing is credible high-end training acceleration.
Gold said that Intel “is really the only player that has the potential to offer a complete component set of products for various aspects of AI computing.”
Gold said the Habana investment for AI silicon in data centers is a “pricey” move, but a good one considering how vital AI will be for Intel and, really, just about every major chipmaker. “I’m sure this won’t be the last acquisition Intel makes in AI,” he predicted. “Intel is betting big on AI.”
Intel estimates the market for AI silicon to exceed $25 billion by 2024, of which $10 billion will be for AI silicon in the data center. Intel expects $3.5 billion in AI revenue in 2019, up 20% from a year earlier.
Intel said Monday it will assess with input from customers how it will integrate Habana with Nervana chips that Intel already produces. Gold said that the two are actually in two different markets with Habana for high performance training and Nervana more a targeted accelerator system for deploying AI systems that have already been trained or don’t need high performance in training when attached to a PC or small server. “The two are somewhat complementary but in two different segments of the market,” he said.