NVIDIA BlueField-4 raises the DPU's maximum network rate to 800 Gb/s and adds a 64-core Grace CPU, PCIe Gen6 x16 and up to 128 GB of LPDDR5x memory. BlueField-3 remains a different, configurable product family: its specifications vary by card, with eight- and 16-core models and 16 GB or 32 GB of DDR5.
The useful comparison goes beyond port speed. NVIDIA now separates BlueField's infrastructure services, called Scale-In, from the shared inference-context storage built around its broader BlueField-4 platform.
Cover: AI-generated conceptual illustration, not a photograph or product rendering.
BlueField-4 vs BlueField-3 specifications
| Feature | BlueField-3 | BlueField-4 DPU |
|---|---|---|
| Maximum network rate | Up to 400 Gb/s; card and port configuration matter | Up to 800 Gb/s |
| CPU | Eight or 16 Arm A78 cores, depending on model | 64 Arm Neoverse V2 cores in a Grace CPU |
| Memory | 16 GB or 32 GB DDR5, depending on model | Up to 128 GB LPDDR5x |
| Host connection | PCIe Gen5 x16 on listed cards; some support an additional x16 connection | PCIe Gen6 x16 |
| On-board SSD | 128 GB on the listed cards | Up to 512 GB |
The table combines NVIDIA's BlueField-3 card specifications with its BlueField-4 DPU datasheet. These are component capabilities, not measured application throughput.
For BlueField-3, the distinction between cards is material. NVIDIA lists an eight-core B3140H SuperNIC with 16 GB of memory, while P-Series DPU configurations use 16 cores and 32 GB. A B3220's listed Ethernet rate reaches 200 Gb/s; other configurations reach 400 Gb/s. Treating every BlueField-3 card as a 400G, 16-core, 32 GB device would hide those differences.
The practical procurement question is therefore which ordering part number and configuration a server actually contains. A generation name alone does not establish the port layout, memory capacity or available host connections.
What the advertised sixfold compute gain means
NVIDIA's Vera Rubin architecture comparison lists six times the DPU compute performance for BlueField-4 relative to BlueField-3. It also compares 400 with 800 Gb/s of network bandwidth and 32 with 128 GB of memory capacity in its selected baseline.
Those are NVIDIA's component comparisons. They do not establish six times the throughput for an inference model, a storage appliance or an entire GPU cluster. The deployment still needs a workload-specific test of the services being offloaded and the surrounding network.
Scale-In connects and manages the infrastructure around compute
In NVIDIA's August Scale-In explanation, BlueField-4 runs infrastructure services outside the tenant host. DOCA supplies the software and microservices, and Spectrum-X Ethernet connects the access and storage paths.
This includes security, tenant isolation, provisioning, telemetry and data access. In the described Vera Rubin NVL72 system, ConnectX-9 SuperNICs carry scale-out workload traffic, while BlueField-4 handles the services that connect, protect and manage the server. The two components have distinct jobs even though the DPU itself incorporates ConnectX-9 networking technology.
For the earlier rack generation, our GB300 NVL72 networking guide maps NVLink, ConnectX-8, BlueField-3 and out-of-band management. That fabric map is a better starting point than assuming a newer DPU replaces every network path.
KV-cache storage is a separate deployment role
NVIDIA's CMX architecture explanation describes an Ethernet-attached flash tier shared across an inference pod. It calls this G3.5: a context-cache layer between local memory/storage and general-purpose shared storage.
Dynamo and NIXL coordinate context movement, including staging reusable KV blocks back into host or GPU memory before decoding needs them. This is an extension of the memory hierarchy; it does not make flash equivalent to GPU HBM or remove active generation's need for GPU memory.
There is also a product distinction. The DPU datasheet describes the BlueField-4 DPU with a Grace CPU, and separately identifies the BlueField-4 STX storage processor with a Vera CPU and ConnectX-9 SuperNIC. STX supports CMX and other AI-storage workloads. Buyers should keep those processor specifications separate when reading platform-level announcements.
Questions to resolve before sizing a deployment
Start with the exact DPU or storage-processor configuration. Then identify whether the bottleneck is host infrastructure processing, network transport, storage access or KV-cache reuse. Those are different reasons to buy hardware.
Ask the supplier to specify supported DOCA features and firmware, demonstrate the intended service mix, and test the end-to-end path. The BlueField-4 datasheet explicitly directs readers to DOCA release notes for feature availability. Hardware capability and enabled software support need to be checked together.
Sources
This article was researched and fact-checked against the following sources:
- NVIDIA BlueField-4 DPU Datasheet (resources.nvidia.com)
- NVIDIA BlueField-3 Networking Platform: Card Specifications (networking-docs.nvidia.com)
- Inside the NVIDIA Vera Rubin Platform (developer.nvidia.com)
- NVIDIA BlueField-4 Powers New Scale-In Network Infrastructure for Agentic AI Factories (developer.nvidia.com)
- Introducing NVIDIA BlueField-4-Powered CMX Context Memory Storage Platform (developer.nvidia.com)
