AMD: Pensando gives us better-than-AWS networking tech to rule the cloud
Equip Amazon rivals with flexible networking chips? Sounds Epyc
With cloud service providers like Amazon Web Services increasingly building their own silicon, it's no surprise that large chip companies like Intel and AMD are trying to find new ways not just to remain relevant in the cloud but to become can't-miss vendors for next-generation architectures.
That was top of mind when AMD made its $1.9 billion bid for networking chip provider Pensando, announced on Monday, top executives from both companies told The Register. And they didn't shy away from their belief that Pensando's networking tech is superior to AWS' custom Nitro silicon.
With this, AMD hopes to arm AWS' rivals and any firm adopting cloud-like architectures with Pensando in competition with the SmartNIC/DPU/IPU efforts of other chip companies, including Nvidia and Intel.
"The value that candidly AWS has shown with Nitro and offloading that stuff and standardizing platform interfaces underneath was pretty evident," said Forrest Norrod, the head of AMD's Data Center Solutions Business Group. "Pensando was very interesting as we were looking at it in that I really believe they're two to three years ahead of anybody else in the industry."
Pensando calls its core offering the "distributed services platform," and its main components consist of a programmable packet processor — which can be wrapped in a PCIe card for any server — and a comprehensive suite of software. With the hardware and software combined, it can speed up workloads for networking, security and storage in a variety of different form factors, from the edge to the cloud.
Soni Jiandani, co-founder and chief business officer at Pensando, said the platform, which is now onto the second generation of its programmable packet processor, is eight to 13 times faster than AWS' 100gbps Nitro system. No detailed comparisons were provided, unfortunately, but Jiandiani said Pensando's tech already has Microsoft Azure hooting and hollering about its capabilities.
"If you look at Microsoft Azure, they call us a 'limitless networking solution,'" she said. "They're embracing our entire stack, and [we're] delivering to them 40 times the connections per second increase compared to where they were today. And they are leapfrogging Amazon, the single largest competitor, by a factor of three when they go into production with us."
Microsoft Azure is among a handful of customers that praised Pensando in canned statements provided in AMD's Monday press release about the deal. The other companies piping up about Pensando were not insignificant: Oracle Cloud, Hewlett Packard Enterprise and Goldman Sachs. Another major customer cited was IBM Cloud, and AMD said all these companies have deployed Pensando's tech "at scale."
Jiandani said this, plus buy-in from other unnamed customers, has translated into 100,000 units deployed in the real world. It's a relatively small number, but Jiandani said Pensando expects to "grow that immensely" once it becomes a part of AMD when the deal closes, which is expected to happen by the end of June.
More than just SmartNICs: Pensando's versatility a big draw
Norrod said the fact that Pensando already has so much silicon in production, supported by solid software, was a major draw for AMD. But the five-year-old chip startup was also attractive because companies can use Pensando's networking technology for more than just standard SmartNICs.
While Oracle Cloud is using Pensando's chips for SmartNICs, other companies have expanded into other use cases, according to Norrod. This includes HPE's Aruba networking division, which is using Pensando for smart switches with built-in firewall security; NetApp, which is using Pensando to speed up storage workloads in new appliances; and other cloud customers who are using Pensando to offload software-defined networking and storage workloads from CPUs.
"We started out thinking this mostly as 'we need to fill the SmartNIC hole.' The application of the technology is way beyond that," he said.
- AMD to acquire SmartNIC and networking firm Pensando for $1.9 billion
- SmartNICs, IPUs, DPUs de-hyped: Why and how cloud giants are offloading work from server CPUs
- ESXi-on-Arm is real and VMware will use it to run networks, storage, and security on SmartNICs
- HPE goes on the warpath, attacks AWS over vendor lock-in
What makes Pensando so versatile is that its chip architecture can support several simultaneous workloads at high speed, Norrod said, which is a major reason why AMD sees Pensando's networking tech peacefully co-existing with the Alveo SmartNIC portfolio from its recently closed Xilinx acquisition. The main difference is that Alveo SmartNICs can offer much higher performance, but they can't support as many workloads at the same time, and their FPGA-based nature means more heavy lifting.
"Pensando offers essentially software programmability that [provides] a great deal of flexibility, and they provide great performance across a wide range of simultaneous workloads or simultaneous data flows," he said.
"But with Alveo, we can hit absolute maximum performance with essentially hardware implementation of that whole pipeline, albeit at the cost of FPGA programming and probably with less simultaneous features supported or less simultaneous flows supported."
So rather than making the Pensando and Alveo teams compete for the same business, Norrod said, he sees the two technologies as complimentary and key to building "the best possible networking technology franchise." To support this view, the executive cited an unnamed company that is using both Pensando and Alveo technologies concurrently.
"They have a very clean view of where one fits and the other fits, so we think doing this gives us the broadest possible range to address any requirement in this market," he said.
AMD may build chiplet designs using Pensando and other silicon
One of the more intriguing takeaways from our interview is that AMD is likely looking to integrate Pensando's chips onto silicon with other chip technologies in its portfolio. This is not a complete surprise since we are already seeing this happen with Nvidia integrating SmartNICs and GPUs together in some of its BlueField DPUs. Pensando's own silicon already incorporates Arm CPU cores.
Norrod said AMD is "absolutely" exploring the possibilities of combining Pensando with the company's other chip technologies as he sees Pensando as "applicable in a lot of places."
"If you look at where things are going around disaggregated or shared infrastructure, I think there's a lot of really interesting things we can do by blurring the lines of existing system logic, CPU, SmartNIC, I/O sub-subsystem boundaries," he said. "I think there's a lot of interesting blurring that can be done to really drive efficiency and performance up simultaneously, so stay tuned."
Norrod said there are a few main factors at play that would allow AMD to mix-and-match different silicon from its portfolio: the rise of the Compute Express Link interconnect standard that will enable different components to share memory, the increasing performance of networking technologies provided by companies like Pensando, and AMD's experience of making designs with chiplets, which the company has used for more recent generations of its Ryzen and Epyc CPUs.
"It gives you the ability to reformulate systems much more easily than if you have to deploy things in a monolithic fashion, so it really opens up the canvas of what you can do, where you can put compute, where you can put these various elements [in a single design]," he said of chiplets.
But regardless of AMD's potentially far-out chiplet dreams for Pensando, Norrod said the startup's networking tech fills an important gap in AMD's portfolio that complements its CPUs, its GPUs, and the "adaptive," FPGA-based chips from its Xilinx acquisition. This, he added, will allow AMD to provide a "complete solution" for high-performance data centers and better compete for more cloud business.
We asked Norrod if AMD thinks it can someday convince AWS to adopt Pensando's tech over the cloud provider's homegrown Nitro chips — which seems fairly unlikely from our perspective, at least for now — and he said he has his eye on the larger data center market, not just one company.
"We're going to try to produce the most compelling leadership technology we possibly can and offer it to any customers that want to embrace it," he said.
"They're all going to make their own decisions based on a number of internal factors. I don't want to say we're targeting any one particular guy. We're just trying to build the most compelling roadmap conceivable." ®
- AWS Graviton
- Black Hole
- Cloud native
- Content delivery network
- Edge Computing
- Google Cloud Platform
- G Suite
- Hybrid Cloud
- Network switch
- Private Cloud
- Public Cloud
- Radio Access Network
- Software-defined network
- Streaming video
- Submarine cable
- Systems Approach
- World Wide Web