How Standards Are Unleashing the Power of DPUs for Cloud Computing

The slowing of Moore’s Law is driving a new approach to compute as single core performance is flattening out. To handle the increasing demands of data-centric workloads, hyperscalers and modern cloud data centers are looking for a new class of programmable processors that can efficiently process and move data at scale. They also need to support growing deployments of advanced applications – including Generative AI models – that require GPUs, faster networking and distributed storage. Today’s cloud infrastructure is custom built, from SSDs to HDDs, SmartNICs to video accelerators, and the last standardized component, the server CPU, will not cut it as a universal general-purpose processor moving forward. Enter data processing units, aka DPUs.

What is a DPU?

A DPU, like a CPU , includes a programmable processor, but it also has custom-built accelerators to handle data-centric workloads such as networking, security, storage, and analytics. DPUs enhance the efficiency and performance of data centers by offloading those data-centric workloads from the CPUs, liberating resources for application processing. They can be scaled to accommodate workloads increasing in volume and complexity as data center needs grow and become more intensive. Additionally, DPUs can be added to existing hardware infrastructure, allowing for a flexible and adaptable data center architecture.

Most DPUs employ Arm cores as the CPU component, capitalizing on the power efficiency and scalability of the Arm architecture. Examples include NVIDIA’s BlueField, Marvell’s OCTEON, Intel’s Mount Evans IPU (a variant of DPU), and Amazon Nitro System.

Although introduced in the early 2010s, DPUs have now gained popularity (partly due to the exponential growth of data and networking demands) among cloud data centers and hyperscalers. But for DPUs to truly be adopted at scale, they must be embraced and be made easily accessible to the software community. Ultimately, it’s the software that drives the value of any product in today’s environment.

To enable a “software just works” experience for DPUs and other devices, we need industry standards that define system and firmware specifications for how hardware should behave for OSes and hypervisors to just work and interoperate.

How Are Standards Enabling DPUs?

Arm offers software and hardware solutions to facilitate DPUs for cloud computing. Arm SystemReady implements a compliance and certification program that stipulates system and firmware specifications for how hardware should behave for OSes and hypervisors in order to just work and interoperate. Arm also collaborates closely with industry standards such as Open Programmable Infrastructure project and Open Compute Project to address areas that enable a “software that just works” experience for data center solutions.

This benefits the entire value chain of the ecosystem, from hyperscalers, silicon vendors, OEMs, ODMs, OSVs, and ISVs to developers.

For silicon vendors – there are significantly reduced software development costs and faster time-to-market. There is no need for silicon vendors to maintain custom versions of a kernel, drivers, or even of a mini-distribution. One cannot emphasize this enough – programming for a standard Arm processor significantly streamlines the software development effort.

And for device manufacturers and developers, they get a lot more value out of these standards-based DPUs such as:

• The software deployment model is streamlined, which enables developers to depend on upstream distributions and receive security updates and bug fixes.
• The developer community can be expanded, both in the open-source domain and among commercial software vendors.
• The adoption of modern methodologies, such as cloud native, DevOps, CI/CD, containerization, etc., is facilitated. These methodologies can pose significant challenges when the silicon is constrained by a specific version of a specific distribution.
• The market scalability is enhanced, as the silicon can be applied to a broader range of markets and applications, without being restricted by a particular software distribution.
• The customer on-boarding process is smoother. The shelf-life is longer, and the system extension with additional devices is easier, due to higher interoperability.

“Our strategic collaboration with Arm has helped Marvell drive and accelerate the adoption of industry-leading DPUs, like the OCTEON 10 family of processors, for 5G infrastructure, cloud data center and enterprise markets.” said Cary Ussery, Vice President of Software and Support at Marvell. “Standards like Arm SystemReady enable consistent hardware behavior for OSes and hypervisors in a way that facilitates software reuse and platform scalability from core to edge.”

“As data centers transform into AI factories, networking infrastructure must evolve to keep up with the spiking efficiency and scalability demands,” said John Kim, Director of Networking at NVIDIA. “NVIDIA BlueField DPUs and the DOCA software framework are enabled with Arm architecture to offload, accelerate, and isolate advanced networking, storage, and security services – massively boosting data center efficiency and scale.”

In summary, Arm is uniquely positioned to empower our partner ecosystem to lead the world’s computing infrastructure with standards-based solutions that unleash the power of DPUs for cloud computing.

Get started with Arm SystemReady