As the AI boom puts different requirements on the hardware needed to serve modern enterprise workloads, data processing units (DPUs) are drawing interest. The microprocessor pioneered by hyperscalers and cloud companies might soon make sense for on-premises enterprise workloads.
Nvidia launched the first of its BlueField DPUs, the BlueField-2, in 2020. At the time, the idea was to allow cloud providers to offload the cloud control plane onto the DPU to free up the host CPU, said Itay Ozery, director of product marketing for networking at Nvidia.
That allowed the hyperscalers to offer bare-metal computing as a service. Multitenant cloud services need a hypervisor to function, but the hypervisor is stripped out in bare-metal computing to eliminate virtualisation overheads, which is particularly compelling in the era of AI workloads and power-hungry applications. Instead, DPUs allows for multiple tenants by offloading network and storage virtualisation and isolating tenant workloads.
“It's been widely used in the hyperscalers and is slowly coming into the enterprise,” said Manosiz Bhattacharyya, chief technology officer at cloud computing firm Nutanix. “Companies interested in some of the higher-end solutions, specifically for AI and high-intensive compute, who actually want all the cores of your host to be used by the application, would want to offload a lot of that.”
Down to the core
DPUs have two core components, a network card and ARM cores, which can be customized by the chip provider, Bhattacharyya said. The BlueField-4 has a 64-core Grace CPU, with throughput of 800 Gb/s, doubling from 400 Gb/s in the previous generation, allowing data-center operators to offload network virtualization workloads, which can take up as much as 20-25% of the host core, Bhattacharyya said.
