Intel continues to make larger bets on the market for AI chips and now has its first customer in IBM Cloud which will be using Intel’s Gaudi 3 that is built specifically as an AI accelerator chip. IBM and Intel on Thursday announced that Gaudi 3 will be available early next year via IBM Cloud. The new acclerators will work in hybrid and on-premise environments, with IBM integrating Gaudi 3 support into its Watsonx AI and data platform.
Justin Hotard, general manager of Intel’s data center and AI division said: “The full potential of implementing AI to solve real-world problems is realized when the solutions are accessible, adoptable by users across geographies. IBM also announced that starting in the second quarter of 2021, Gaudi 3 AI accelerators and Xeon CPUs will be integrated with IBM Cloud to provide new AI capabilities, catering an increasing need for secure next-generation artificial intelligence computing solutions at scale.
Intel introduced its new Gaudi 3 chip last month as part of the company’s efforts to take on industry leaders Nvidia and AMD in AI chips. For the third and last of its Gaudi lineup—the series originally came about after Intel spent roughly $2 billion to buy out AI startup Habana Labs in 2019—Gaudi 3 is designed to provide performance that can keep up with pricing unlike any other.
Earleir this year, Intel demoed Gaudi 3 reference designs which will allow partners (Lenovo certainly one of them) to slot the chips into their own servers. These designs will feature Ethernet connectivity that Graphcore claims to have a key IP on, which it aims against Nvidia’s InfiniBand technology, alongside the pairing of Gaudi 3 with Intel Xeon 6 processors.
Still, Gaudi 3 is launching into a difficult environment for Intel. Though the chip certainly ticks a lot of boxes for folk that might already be locked into Nvidia, Intel has its work cut out to win them around. That’s tiny compared against the $3.5 billion AMD believes it can get from its state-of-the-art (MCM-based) Instinct MI200-series GPUs or even, more dramatically still, Nvidia with an estimated 40-plus-billion-dollar-per-year data center business; of which NVIDIA-Grace and OAM are but a small part. While Gaudi 3 may show good benchmarking results, making this a success in the market will still be Intel’s biggest obstacle.
Intel’s CTO, Greg Lavender, went so far as to say in July that Intel could get the number two AI chip-maker spot behind industry leader Nvidia. Yet not long after this gleeful pricing announcement, Intel served up a $1.6 billion quarterly loss in Q2 — only to follow it by announcing 15,000 layoffs and more than $10 billion in cost-cuts targeting 2025.
And as if Intel’s problems weren’t enough, next-gen AI chip Blackwell is expected from Nvidia in Q4 following a short manufacturing hiccup. This forms a to four-time improvement over Nvidia’s H100, which will no doubt have this being marketed by Intel against its former performance target.
Complicating the matter, Intel has also declined to single out how Gaudi 3 stacks up versus Blackwell in expectations of such chips not being available until they’re on sale. With Nvidia poised to push the boundaries of AI performance again, Intel finds itself at a critical moment in trying to remain competitive with an increasingly agile tech giant.
