What does a DPU do?
A data processing unit is a hardware accelerator specifically geared towards data processing and data-centric computing. It differs from other hardware accelerators such as CPU and GPU in that it exhibits an increased degree of parallelism and MIMD architecture.
Many organizations use the DPU for supercomputing tasks such as AI and big data. To decide if your organization needs a DPU in its data center, understand its use cases and drawbacks.
What does a DPU do?
A DPU offloads networking and communication workloads from the processor, allowing the processor to tackle application support tasks instead. It focuses on data-centric workloads such as data transfer, data reduction, data security, and analytics. The chip features a specialized design that combines processor cores with hardware accelerator blocks. This design makes the DPU a more versatile general purpose chip than the GPU. The DPU has its own dedicated operating system, which means that you can combine its resources with the resources of your main operating system and that it can perform functions such as encryption, erasure coding, and compression or decompression.
Cloud and hyperscale providers were the first to adopt this technology. However, vendors like VMware have started adding support for DPUs to their offerings, making them more appealing to other organizations.
Support for storage with a DPU
Due to the versatility of DPUs as a processing unit, you can use DPUs to support storage in your data center. For example, you can speed up access to NVMe storage devices by connecting them to the PCIe bus of the DPU.
DPU also gives you better access to remote storage devices that rely on NVMe-oF. The DPU presents these remote storage devices to the system as standard NVMe devices. This optimizes your connectivity to remote storage because it means that you no longer need special drivers to connect to these remote storage devices.
DPU and data-centric architecture
The DPU is only part of a data-centric architecture. This paradigm forces you to build an infrastructure around data requirements, instead of forcing data to fit the infrastructure. It makes data the primary consideration for application development, business decisions and infrastructure deployment. A data-centric organization treats data as its core asset, eliminates silos, and mitigates sprawl by implementing a single data strategy for multiple applications.
Data-centric hardware, such as the DPU, facilitates the movement and delivery of data. It must offer high availability and reliability, and must allow the entire organization to access shared data in real time. Its performance, capacity, scalability and security must evolve to meet new workload demands and adapt to new technologies.
In the context of a data-centric architecture, the DPU solves the inefficiency of server nodes when it comes to data-centric computing, and it also solves slow or inefficient data transfer or sharing between server nodes. .
The growing popularity of the DPU
In 2020, the startup Fungible released the first version of the DPU. He created two separate versions of the processing unit: one for storage and one for networking. Both versions of the Fungible DPU included memory and on-chip processing for tasks such as storage, security, networking and virtualization. Fungible designed them to provide the benefits of hyperconverged infrastructure, but with better sharing of storage and networking resources.
Since the release of Fungible’s DPU, vendors such as Nvidia and Intel have released their own versions of this technology. In June 2021, Intel released its infrastructure processing unit chip, which does the same job as a DPU. In the wake of Intel, in July 2021, Nvidia unveiled its own DPU. These processing units, along with those from additional competitors such as Marvell and AWS, offload all host processor tasks to speed up and streamline data compute workloads. Nvidia expects carriers and cloud providers to embrace its technology first, but the boom in DPU offerings from major vendors means you might see it in other data centers soon.