NVIDIA classifies the DPU (data processing unit) as a new programmable processor category joining the CPU and GPU as a third pillar of computing, built to offload networking, storage, and security tasks from host processors. NVIDIA's BlueField-3 runs these functions at 400 Gb/s line rate, AMD's Pensando Giglio and Elba DPUs each deliver dual 200 Gbps line-rate acceleration, and Intel's IPU targets the same infrastructure-offload function to raise CPU utilization.
What Is a DPU, and Why Does NVIDIA Call It Computing's Third Pillar?
NVIDIA defines the DPU as a new class of programmable processor set to join the CPU and GPU as one of three pillars of computingCITE:E1. NVIDIA further separates the three roles: the CPU is for general-purpose computing, the GPU is for accelerated computing, and the DPU moves data around the data center and performs data processingCITE:E2.
How Does a DPU Offload Infrastructure Tasks Such as Networking, Storage, and Security From the CPU?
NVIDIA states that its BlueField DPU line offloads networking, storage, and security services from host CPUs to keep on-premises AI infrastructure runningCITE:E3. Intel describes its IPU as built for the same purpose, calling it infrastructure task offload to improve CPU utilization and performanceCITE:E7.
How Does NVIDIA's BlueField-3 DPU Deliver Line-Rate Infrastructure Processing?
NVIDIA specifies the BlueField-3 DPU as a 400 Gb/s infrastructure compute platform with line-rate processing of software-defined networking, storage, and cybersecurityCITE:E4. This throughput figure sits on top of the offload role NVIDIA already assigns to the BlueField line for networking, storage, and securityCITE:E3.
What Are the Specs of AMD's Pensando Giglio and Elba DPUs?
AMD states that its Pensando Giglio DPU, built on a second-generation architecture with enhanced power and performance efficiency, delivers dual 200 Gbps line-rate acceleration for data center networking, storage, and securityCITE:E5. AMD also describes the second-generation Pensando Elba DPU as fully P4 programmable and optimized for high throughput, enabling networking, storage, and security services at the same dual 200 Gbps line rateCITE:E6.
How Do NVIDIA and AMD's DPU Line-Rate Specifications Compare?
NVIDIA's BlueField-3 platform is rated at 400 Gb/s line rate, while AMD's Pensando Giglio and Elba DPUs are each rated at dual 200 Gbps line rateCITE:E4CITE:E5CITE:E6.
| Vendor | Product | Line-Rate Spec | Function Scope |
|---|
| NVIDIA | BlueField-3 DPU | 400 Gb/s | Networking, storage, cybersecurityCITE:E4 |
| AMD | Pensando Giglio DPU (2nd-gen architecture) | Dual 200 Gbps | Networking, storage, securityCITE:E5 |
| AMD | Pensando Elba DPU (2nd generation) | Dual 200 Gbps | Networking, storage, securityCITE:E6 |
What this means: Across NVIDIA, AMD, and Intel product pages, the same functional definition recurs — moving networking, storage, and security processing off the host CPUCITE:E3CITE:E7. NVIDIA frames its offload capability with a single unified throughput figure, 400 Gb/sCITE:E4, while AMD frames the same function across two DPU generations with an identical dual 200 Gbps specCITE:E5CITE:E6 — meaning the stated throughput held constant from Giglio to Elba even as AMD labels Elba a second-generation, fully P4-programmable productCITE:E6.
Author's Take・林紀旭 James Lin
The three vendors' spec sheets converge on the same architecture story but diverge on how they express it. NVIDIA anchors BlueField-3 to a single 400 Gb/s line-rate figure, while AMD pins both its second-generation Giglio and its second-generation Elba to an identical dual 200 Gbps spec. That AMD held the same throughput number across two DPU generations, while changing programmability — Elba is fully P4-programmable — rather than raw bandwidth, suggests AMD's current generational bet is on flexibility, not top-line speed. NVIDIA and Intel, for their part, both describe their DPU and IPU lines as solving the identical problem: freeing host CPU cycles by offloading networking, storage, and security. The metric worth watching next is whether AMD's following Pensando generation finally moves off the 200 Gbps-per-port mark, which would signal a shift back toward bandwidth competition with NVIDIA's single-figure approach.