A year ago, AMD plunked down $1.9 billion to acquire privately held data processing unit (DPU) vendor Pensando. A year later, what has AMD been doing with Pensando and is the big bet paying off?

To answer that question SDxCentral talked to Soni Jiandani, who currently serves as the corporate vice president of the Networking Solutions Group at AMD. Jiandani was part of the team of ex-Cisco employees that started Pensando in 2019. She noted that over the past year, the Pensando business unit within AMD has grown, advancing the way that DPU technology can be integrated into smart switches and infrastructure acceleration systems. A DPU is a purpose-built piece of silicon technology that is designed for network data traffic. The basic idea is that by offloading data traffic to a DPU, compute intensive tasks can be optimized on CPU and GPU resources.

"If you look at the journey that AMD is on in the data center, it's really to look at how to go about building the industry's highest performing adaptive computing portfolio,"Jiandani said. "I think DPUs are the most efficient way to deal with things that need to have data processing at the edge."

With the rise of microservices-based architectures, there is more east-west networking traffic inside of data centers than ever before. At the same time, Jiandani said that there is an increasing amount of data processing occurring at the edge of the network.

[caption id="attachment_128264" align="alignnone" width="674"] "DPUs are the most efficient way to deal with things that need to have data processing at the edge." -- Soni Jiandan, corporate vice president of the Networking Solutions Group at AMD[/caption]

Jiandani noted that she has seen reports that have estimated the overhead of running data infrastructure services to be anywhere from 31 to 83%. In her view, that means at a minimum, a third of all the CPUs today are being wasted just to run infrastructure services.

"You buy CPUs to run workloads and applications," she said. "For cloud providers to put aside 31% of infrastructure means they can't generate revenue for those CPUs, nor do they help the enterprise either."

The need for DPU type technology is seen by more than just AMD. In June 2022, the Linux Foundation launched the Open Programmable Infrastructure (OPI) Project as an effort to create open source standards for DPUs as well Infrastructure Programming Units (IPUs). AMD supports the effort as do its primary rivals in the space which include Intel and Nvidia.

"By actively participating and contributing in standards bodies like OPI, we believe it further accelerates the ability for customers to embrace the architecture in a more comprehensive manner and bring along an ecosystem to deliver the entire solution," Jiandani said. "It allows the vendors to work and interoperate with each other in a much more seamless manner."

As part of AMD's networking group, Jiandani said that her group is taking a system approach to DPUs and how they can help cloud operators and enterprises alike to optimize compute and infrastructure resources.

The systems approach includes a software defined networking approach that supports the P4 programming language. The combined hardware and software system from AMD can enable both data traffic offload as well as better observability and security, according to Jiandani.

It's also an approach that has different form factors. The AMD Pensando system approach can be used to enable a smart switch, which is how HPE's Aruba networking business unit uses the technology. It can also sit directly inside a server that can be deployed at the edge or in the cloud.

"We are firmly form factor agnostic, depending on the partners and the customers and the way they are ingesting the technology," she said. "Ultimately, the building blocks of the software stack and the DPU are the same."

In terms of new hardware, Jiandani hinted that there will be more to come from AMD Pensando in 2023, though she revealed few details. Overall she emphasized that when it comes to DPU networking gear, product innovation is about evolutionary steps forward, not revolutionary.

"Today when we are selling our system to a cloud customer, that cloud customer wants to see a lifespan of seven years behind that architecture," Jiandani said. "That's why it's very important to not think like a chip vendor, but to think like a systems vendor, which is the single most important thing today that differentiates us."