This time last year, AMD had no networking products. Today, it has three, thanks to two separate acquisitions. It's a shift that's strengthening the vendor's position against competitors like Intel and Nvidia by providing a full suite of silicon, including for the enterprise.
To fill out its portfolio and stay competitive, AMD had to build up its offerings of workload accelerators and its networking technologies. It's a very active market, with Nvidia (BlueField), Intel (FPGA-based smartNICs), Marvell Technology (Octeon), and Broadcom (Stingray) all competing for the smartNIC market. AMD risked being left behind.
“Right now, Intel can walk into customers and offer complete solutions, which include intelligent NICs,” says Brandon Hoff, research director for enabling technologies, networking and communications, at IDC. “This is a really strong move for AMD to be able to really have a better portfolio to offer their customers.”
The ability to offer compute and networking capabilities makes AMD “a one-stop shop for both compute and networking,” he says.
Buying Xilinx and Pensando
AMD's first significant step in this direction was the acquisition of Xilinx in February 2022. Known primarily for its Alveo FPGA adaptive accelerator chips, Xilinx also had multiple networking cards, including ultra-low latency networking cards it acquired when it bought SolarFlare in 2019.
The second step was the acquisition of Pensando, a start-up AMD bought in May 2022 for $1.9 billion. The Pensando Distributed Services Card has a data processing unit (DPU) designed to offload networking work from the CPU, freeing up the CPU to do data processing while the card handles network traffic.
Despite having this wealth of networking cards, AMD doesn’t have to deal with overlap; the SolarFlare, Alveo, and Pensando product lines compete in separate markets.
Forrest Norrod, senior vice president and general manager of the Data Center Solutions Business Group at AMD, called the SolarFlare line “the de facto standard for high-frequency [stock] trading for anybody doing ultra-low-latency networking in financial services.” He says SolarFlare has well over 50 per cent of that market segment.
AMD thinks the technology behind those cards can be applied to other areas where low-latency Ethernet is important, "for example, the fabric for HPC, or for the back-side fabric applications for GPUs used for AI training,” according to Norrod.
The Alveo line is used by both hyperscale and enterprise customers in a variety of environments, most commonly networking, compute, AI/ML inference acceleration, video and image processing, financial computing, computational storage, and video analytics, according to AMD.
Pensando gear provides DPU and distributed-services hardware and software that fit a broader set of customers who need them to accelerate infrastructure tasks such as networking, security, and storage in their data centres.
In addition to the technology matches among AMD and its acquisitions, they share similar corporate cultures, Norrod says. “And that's an under appreciated, aspect, I think. If you try to integrate something that has an antithetical culture, it's going to blow up. You're never going to integrate it,” he says.
Processors for enterprises, cloud, and hyperscalers
Over time, customers can expect a broad set of tightly integrated silicon, says Glenn O’Donnell, vice president and research director with Forrester Research. “The bigger opportunity is increasing its market share in its current business by offering this broader family,” he says.
IDC’s Hoff says AMD will be able to offer a complete range of data centre processing without relying on other vendors for the networking piece, including smartNICs that can offload networking chores from CPUs. “The different networking products give AMD more flexibility to offer enterprise customers cloud/hyperscale-class networking,” he says.
And AMD can provide better integration of CPUs, GPUs, and DPUs without having to rely on third parties. “That gives AMD even more flexibility to offer enterprise customers cloud/hyperscale-class networking in the server,” Hoff says.
“I think [customers] will see that we are going to continue to scale up,” Norrod says. “Pensando has capabilities to support more customers, particularly on the enterprise side.” At the recent VMware Explore conference, AMD announced an 800G Pensando processor (the current chip is 400G) for 2024.
Norrod also predicts that with AMD smartNIC and smart-switch technology, enterprises can acquire efficiency on par with cloud providers. He says the Pensando gear is similar to AWS’s Nitro DPU, which offloads network traffic processing from the CPU.
“Amazon with Nitro was a great example of what you can do through smart deployment of a smartNIC in the data centre… freeing up a lot of compute cycles and making their cloud computing much more efficient,” he says.
AMD plans to expand the use of all three of its networking products, but it will take a while for the technology to make it through the product pipeline, according to Norrod. “You'll start to see [advances] show up in products in about two to three years.”