Intel executives outlined the company’s data-center strategy for the next year Tuesday, mapping out a four-year plan for its line of infrastructure processors and several new iterations of existing AI chips.
Intel’s foray into what it calls infrastructure processing units (IPUs) follows other efforts by rivals such as Nvidia and AWS: Both recognized that the demands of modern computing require that networking-related tasks run on a dedicated engine that delivers performance beyond what is offered by a CPU or GPU.
For years, cloud computing providers were simply able to buy ever-larger numbers of powerful graphics and processing chips that got more powerful every couple of years. But as it has become more difficult to cram smaller features onto silicon, server designers have looked for other ways to improve performance.
“IPU is a key part of the future data center architecture,” Intel’s head of Ethernet Products Patty Kummrow said in a briefing with reporters ahead of a launch event Tuesday in Dallas, Texas. “We have talked about our data center of the future, [and] we see the IPU as a critical piece to enable all those optimizations and performance performance drivers that our customers see.”
While Nvidia introduced its own data-processing unit in 2020, Intel didn’t immediately embrace the idea, and launched the first iteration of the IPUs last year. Intel designed its first batch of IPUs with Google, and Kummrow cited demand for the chips from financial institutions because of their security requirements. Separating the infrastructure operations from the core computing has made both tasks significantly more efficient, she said.
“We're seeing a lot of demand and applicability for these devices, even beyond the hyperscale data centers all the way out to the edge,” Kummrow said.
That initial IPU project went well enough to prompt Intel to commit to building four generations of IPUs through 2026, and a set of software tools to help run them. The planned products will scale in speed and complexity, and be driven by data-center operator demands for networking performance, Kummrow said.
The Intel IPUs come in two flavors: one that features a programmable chip that customers can update as they see fit, and a second version with a design that’s locked in, created around the purpose-built chip known as an ASIC that Intel designed with Google. Kummrow promised more details about subsequent iterations of the IPU in the future.
“Security and storage are really emerging workloads that are very, very important, and we're really committed to enabling broad adoption for our customers in the cloud service providers, enterprise and beyond that, all the way to the edge,” Kummrow said.
The addition of several generations of IPUs to Intel’s existing portfolio of data-center chips is another sign of CEO Pat Gelsinger’s influence on the company and his multiyear plan to remake the business.
There are two types of new Intel IPUs.Photo: Intel
Prior to Gelsinger’s appointment to the top boss spot, Intel insisted that its CPUs were sufficiently powerful to meet the needs of the modern data-center customer. Committing to at least four years of IPUs indicates the company’s increasing willingness to admit its processors aren’t enough on their own for today’s data centers.
Separately Tuesday, Intel’s Habana AI unit said that it was releasing a new version of its Goya inference and Gaudi training chips. Intel plans to make them with its seven-nanometer manufacturing tech, which will allow Habana to significantly increase their performance by bolstering the subsystems in the accelerators, including adding ethernet integration onto the devices, among other improvements.
“With Gaudi 2, we are leaping all the way to seven nanometer, and we use that to really upgrade all the major subsystems inside that accelerator,” Habana COO Eitan Medina said.
Intel also said it planned to launch a new generation of data-center graphics chips in the third quarter of this year, which are called Arctic Sound-M and include a hardware video encoder.
Intel’s overall data-center business grew at a healthy clip in the first quarter, with revenue rising 22% to $6 billion — though it missed Wall Street expectations. Despite the strong growth and a strong overall market, however, executives noted that supply issues continue to hamper Intel’s ability to fulfill all the orders coming in from cloud computing companies. The company also continues to lose share to rivals such as AMD, according to Jefferies.
