Data centre infrastructure: the emergence of DPUs in the enterprise data centre
Understanding Data Processing Units (DPUs)
A DPU is a processor aimed at offloading infrastructure functions such as network, storage, and security from the CPU and GPUs on servers or network appliances. Typically, a DPU has its own dedicated CPU, memory, acceleration engine, and operating system. Several benefits should be immediately obvious.
Offloading infrastructure functions can provide up to a 30% saving on the host CPU or GPU.
The DPUs own processor is designed for specific infrastructure functions, which makes them more energy efficient.
By running its own OS, a DPU will continue to function if the host OS fails.
A range of DPU form factors exist, but the most common for enterprise will be PCIe, Onboard, and System on a Chip (SoC). With PCIe, the DPU connects directly into the server’s PCIe bus like a standard NIC. Onboard DPUs connect into the server or appliance motherboard to save room on the PCIe bus. SoC is used on switches where the DPU is packaged with the switch silicon to effectively provide a DPU on each switchport.
In high-performance environments, server and network DPUs may work side by side. Server DPUs provide storage acceleration with features like NVMe-over-Fabrics (NVMe-oF), Zero Trust security, and hypervisor offload. Smart switch DPUs deliver network policy such as micro-segmentation, perimeter security, load balancing, NAT and telemetry.
DPU Hardware
The DPU market is complex but can be distilled to a small number of specialist silicon vendors that manufacture DPUs for server OEMs, network OEMs, hyperscalers, and service providers. Smart switching platforms pair DPU silicon with network operating systems and management software, while security and application-delivery ecosystems integrate DPU acceleration for policy enforcement, inspection, and telemetry.
The primary smart-switching platforms use integrated DPU architectures to provide network-services acceleration inside the switching fabric. A further category of DPU-enabled switches is aimed primarily at hyperscaler environments and is often associated with open networking software.
Several generations of DPU silicon are available, with newer designs improving power efficiency, performance, and feature depth. Current enterprise-class DPUs can support line-rate 400G services, while later-generation hardware can offer improved performance for functions such as Layer 3 to 4 filtering or deep packet inspection.
In smart switches, DPUs are integrated directly with the switching silicon. Backplane switching capacity and DPU services capacity are not always the same: switching throughput may run into multiple terabits, while the DPU complex provides a lower but still substantial services-throughput tier. Even so, the architecture can provide a cost-effective alternative to deploying separate network-services appliances. Typical specifications include the following.
A compact 24-port smart-switch design may use four DPUs, around 128G of combined memory, and estimated network-services throughput of around 800G.
A higher-density smart-switch design may use two newer-generation DPUs, around 128G of memory, and similar services throughput.
Another 24-port design may use two DPUs and 64G of memory while offering around 800G of network-services throughput.
A higher-capacity model may double the DPU count and increase memory to around 128G, providing up to 1.6T of network-services throughput.
Different switch designs can deliver similar services throughput with different numbers of DPUs. This is primarily due to how the DPUs are integrated into the switching chip, including the number and speed of internal links, though vendors may also measure and advertise performance differently.
DPU Use Cases
Finally, we will address the age-old CIO and CTO question of “so what?”. Aside from network and storage performance, the two most obvious use cases are Layer 3 to 7 network services and telemetry, provided via a centralised controller.
Modern DPUs provide traditional Layer 3 to 4 stateful firewalling, Layer 4 to 7 deep packet inspection, and application load balancing at a port level within the smart switch. In some cases, this capability can be delivered at a cost comparable to a standard Top of Rack (ToR) switch. This enables the removal of expensive high-end firewalls, load balancers, and micro-segmentation solutions, reducing complexity, cost, and Total Cost of Ownership (TCO).
We should all be aware of the benefits of analytics and telemetry, so I won’t make the case for them here. DPUs can provide flow-based telemetry, stateful connection tracking, and latency measurements. Having this information directly from the DPU in the switch is operationally invaluable and supports customer experience, security, compliance, and reduced TCO.
Some DPUs also support a P4 data pipeline. For anyone unfamiliar with P4, put simply it allows the vendor to re-programme the chip and change how the DPU handles packets. New capabilities can therefore be introduced far more rapidly, improving investment protection.
Over the next two to three years, we expect customers to realise the benefits of DPUs within their data centre environments. Designs will be simplified and costs reduced, while performance and security improve, with the bonus of better analytics and a more sustainable data centre.
