CPU, GPU…..DPU?

Data centers are running out of CPU headroom as virtualization expands from servers into networking and security—and NVIDIA’s BlueField 3 data processing unit (DPU) is positioned as the fix. The core shift is simple: instead of forcing the CPU to handle high-volume packet forwarding, encryption/decryption, and inspection workloads, a DPU acts like a specialized “server inside a server” that takes over those network functions, keeping the CPU focused on general compute.

Virtualization began by consolidating many physical machines into fewer hosts, with each virtual machine sharing CPU and memory resources. That worked—until the same consolidation logic spread to the rest of the stack. Modern data centers increasingly virtualize switches, routers, firewalls, and security appliances using platforms such as VMware NSX. As more networking and cybersecurity functions move into software, the CPU becomes the bottleneck: it must manage many operating systems and also perform tasks it wasn’t built for, especially as network speeds climb from 1 Gbps to 10, 100, and 200 Gbps. The transcript frames this as a mismatch between general-purpose processing and specialized data-plane work like traffic inspection, encryption, and large-scale data movement.

The first step in relieving that pressure is the smart NIC (network interface card), which offloads some networking and security tasks from the CPU. But smart NICs eventually hit limits as workloads grow and more functions get virtualized. That’s where the DPU enters. NVIDIA’s BlueField 3 is described as an SoC (system on a chip) that can sit on a smart NIC form factor, yet runs its own operating system—specifically, ESXi is installed on the DPU in the VMware environment described. In practical terms, firewall and networking software can run on the DPU so network traffic no longer has to traverse the CPU path.

The operational payoff is twofold: performance and mobility. The lab demonstrates that VMware vSphere 8 with the distributed services engine (previously called Project Monterey) supports Universal Pass Through (UPT) for NVIDIA DPUs. That enables vMotion of a VM that uses the DPU as its network interface without breaking connectivity—something the transcript says fails with PCIe pass-through smart NIC setups.

On raw throughput and latency, the lab’s numbers are stark. With a standard NIC using a 25 GbE cluster, bandwidth is reported at 64 megabits per second, latency around 0.273 ms, and 548,000 operations per second. Switching to the BlueField 3 DPU raises bandwidth to 90 megabits per second, latency to about 0.19 ms, and operations per second to 770,000. Crucially, the CPU remains less stressed in the DPU case, implying not just faster networking but better efficiency.

The broader argument is that DPUs change how virtual machines should be designed and how data centers scale. Rather than adding more servers to spread CPU load—an approach that increases space and power use—the DPU adds specialized processing capacity within the existing server footprint. The transcript also notes that NVIDIA’s DPU programmability stack (DOCA) enables customization, and it points to additional white papers for scaling beyond the single workload and 25 GbE test environment.