This time final 12 months, AMD had no networking merchandise. Right now, it has three, thanks to 2 separate acquisitions. It is a shift that is strengthening the corporate’s place towards rivals like Intel and Nvidia by offering a full suite of silicon, together with for the enterprise.
To fill out its portfolio and keep aggressive, AMD needed to construct up its choices of workload accelerators and its networking applied sciences. It is a very lively market, with Nvidia (BlueField), Intel (FPGA-based smartNICs), Marvell Expertise (Octeon), and Broadcom (Stingray) all competing for the smartNIC market. AMD risked being left behind.
“Proper now, Intel can stroll into clients and supply full options, which embody clever NICs,” says Brandon Hoff, analysis director for enabling applied sciences, networking and communications, at IDC. “It is a actually robust transfer for AMD to have the ability to actually have a greater portfolio to supply their clients.”
The power to supply compute and networking capabilities makes AMD “a one-stop store for each compute and networking,” he says.
Shopping for Xilinx and Pensando
AMD’s first vital step on this course was the acquisition of Xilinx in February 2022.
Identified primarily for its Alveo FPGA adaptive accelerator chips, Xilinx additionally had a number of networking playing cards, together with ultra-low latency networking playing cards it acquired when it purchased SolarFlare in 2019.
The second step was the acquisition of Pensando, a startup AMD purchased in Might 2022 for $1.9 billion. The Pensando Distributed Companies Card has an information processing unit (DPU) designed to dump networking work from the CPU, releasing up the CPU to do knowledge processing whereas the cardboard handles community site visitors.
Regardless of having this wealth of networking playing cards, AMD doesn’t must cope with overlap; the SolarFlare, Alveo, and Pensando product strains compete in separate markets.
Forrest Norrod, senior vice chairman and basic supervisor of the Knowledge Heart Options Enterprise Group at AMD, known as the SolarFlare line “the de facto normal for high-frequency [stock] buying and selling for anyone doing ultra-low-latency networking in monetary companies.” He says SolarFlare has nicely over 50% of that market phase.
AMD thinks the expertise behind these playing cards might be utilized to different areas the place low-latency Ethernet is essential, “for instance, the material for HPC, or for the back-side cloth functions for GPUs used for AI coaching,” in accordance with Norrod.
The Alveo line is utilized by each hyperscale and enterprise clients in a wide range of environments, mostly networking, compute, AI/ML inference acceleration, video and picture processing, monetary computing, computational storage, and video analytics, in accordance with AMD.
Pensando gear offers DPU and distributed-services {hardware} and software program that match a broader set of consumers who want them to speed up infrastructure duties reminiscent of networking, safety, and storage of their knowledge facilities.
Along with the expertise matches amongst AMD and its acquisitions, they share comparable company cultures, Norrod says. “And that is an underappreciated, side, I believe. If you happen to attempt to combine one thing that has an antithetical tradition, it may blow up. You are by no means going to combine it,” he says.
Processors for enterprises, cloud, and hyperscalers
Over time, clients can count on a broad set of tightly built-in silicon, says Glenn O’Donnell, vice chairman and analysis director with Forrester Analysis. “The larger alternative is rising its market share in its present enterprise by providing this broader household,” he says.
IDC’s Hoff says AMD will be capable of supply an entire vary of data-center processing with out counting on different distributors for the networking piece, together with smartNICs that may offload networking chores from CPUs. “The totally different networking merchandise give AMD extra flexibility to supply enterprise clients cloud/hyperscale-class networking,” he says.
And AMD can present higher integration of CPUs, GPUs, and DPUs with out having to depend on third events. “That offers AMD much more flexibility to supply enterprise clients cloud/hyperscale-class networking within the server,” Hoff says.
“I believe [customers] will see that we’re going to proceed to scale up,” Norrod says. “Pensando has capabilities to help extra clients, significantly on the enterprise facet.” On the latest VMware Discover convention, AMD introduced an 800G Pensando processor (the present chip is 400G) for 2024.
Norrod additionally predicts that with AMD smartNIC and smart-switch expertise, enterprises can purchase effectivity on par with cloud suppliers. He says the Pensando gear is just like AWS’s Nitro DPU, which offloads community site visitors processing from the CPU. “Amazon with Nitro was an incredible instance of what you are able to do by sensible deployment of a smartNIC within the knowledge middle…releasing up a whole lot of compute cycles and making their cloud computing far more environment friendly,” he says.
AMD plans to increase using all three of its networking merchandise, however it would take some time for the expertise to make it by the product pipeline, in accordance with Norrod. “You will begin to see [advances] present up in merchandise in about two to a few years.”
Copyright © 2022 IDG Communications, Inc.
