NVIDIA BlueField-3 is a data-processing unit (DPU) designed to move data-center infrastructure work—such as networking, storage, security and virtualization—from host CPUs into a combination of Arm cores and dedicated or programmable hardware engines. At Hot Chips 33 in 2021, the architecture was presented as a way to process that infrastructure software closer to the network and storage traffic it handles.
What NVIDIA presented at Hot Chips 33
In an August 19, 2021 preview, NVIDIA said Idan Burstein, a principal architect on the BlueField design team, would detail BlueField-3 at Hot Chips 33. The design combined general-purpose Arm cores with hardware accelerators for infrastructure workloads. NVIDIA’s framing was that virtualization, networking, storage, security and related cloud-native services consume substantial host processing capacity; its 2021 preview claimed these workloads already used a third of server CPU cycles.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Gvdlink NMFP7E20 Optical Multimode Splitter Fiber Cable 5m (16.4ft) MPO12 to 2xMPO12 LSZH OM4 for... | $49.00 | Buy on Amazon |
The architectural aim is not to replace the host CPU. It is to let the DPU take responsibility for selected infrastructure functions so the host can spend more of its resources on applications. NVIDIA CEO Jensen Huang described the need for a processor designed specifically to offload and accelerate data-center infrastructure software.
How the architecture divides the work
Arm cores handle software-defined control
The formal NVIDIA datasheet specifies up to 16 Armv8.2+ Cortex-A78 “Hercules” cores, with 8MB of L2 cache and a 16MB system-level last-level cache. These general-purpose cores provide a programmable environment for infrastructure software and control tasks. NVIDIA’s DOCA software framework is the development and management layer for DPU applications, covering development, compilation, optimization, provisioning, updating and monitoring.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11#1 Best Overall
- The MFP7E20-Nxxx cable for NVIDIA, is a multimode, 4-channel-to-two 2-channel splitter fiber cable. The Multiple Push On, 12 fiber, Angled Polished Connectors (MPO-12/APC) uses 8 active fibers to transmit light and 4 inactive fibers as strength members. The Angled Polished Connector has a 8-degree polished angle to deflect internal optical back reflections from entering the transceivers and distorting the signal quality
- The 4-channel end is inserted into a Twin port OSFP, 800Gb/s transceiver. The 2-channel ends are inserted into two, single-port 400Gb/s OSFP and/or QSFP112 transceivers which with only 2 fibers can output 200G rates. Two splitter fiber cables are used in the twin-port OSFP transceiver enabling four, 2-channel ends to four transceivers.
- The fibers are “crossover”, Type-B cables enable directly attaching two transceivers together and allow the transmit laser fiber on pin 1 to “crosses over” and align with pin 12 of the opposite fiber end transceiver photodetector.
- The typical usecase is linking OSFP switches to in ConnectX-7 network adapters and/or BlueField-3 Data Processing Units (DPUs) in compute and storage servers.
- Rigorous cable production testing ensures best out-of-the-box installation experience, performance, and durability. For NVIDIA’s optical solutions provide short, medium, and long reach scalability for all topologies, utilizing innovative optical technologies to enable high signal integrity and reliability
Datapath engines handle parallel traffic work
A contemporaneous Hot Chips 33 report describes a separate 16-core, 256-thread programmable datapath accelerator and programmable packet-processing pipelines. The distinction matters: the Arm subsystem can run software-defined functions, while specialized or programmable engines process high-throughput packet, storage, cryptographic and timing operations without routing every operation through host software.
The same report says the NIC subsystem is independent of the Arm system and runs its own real-time operating system. That arrangement allows the Arm side to reboot without rebooting the NIC side. In practical terms, this separates parts of the control plane from the data plane; it does not mean that every DPU service or host workload is unaffected by every reset or failure.
Memory and host connectivity
NVIDIA’s datasheet lists dual DDR5 memory controllers rated at 5600MT/s, 16GB of onboard DDR5, and 32 PCIe Gen 5 lanes. The PCIe switch supports bifurcation into as many as 16 downstream ports. These figures describe the published platform envelope; individual card configurations can differ.
BlueField-3 specifications at a glance
| Specification | Published BlueField-3 capability |
|---|---|
| Arm processor | Up to 16 Armv8.2+ A78 Hercules cores |
| Cache | 8MB L2 and 16MB LLC system cache |
| Programmable datapath | 16 cores and 256 threads |
| Memory | Dual DDR5 5600MT/s controllers; 16GB onboard DDR5 |
| Host interface and switching | 32 PCIe Gen 5 lanes; PCIe switch bifurcation supporting up to 16 downstream ports |
| Networking ceiling | Up to 400Gb/s Ethernet or one-port NDR 400Gb/s InfiniBand, depending on configuration |
| Other networking configurations | Dual NDR200 or HDR options on some configurations |
These are datasheet-level maximums and options, not a promise that one card combines every listed feature. NVIDIA’s current platform documentation describes physical PCIe cards with host connectivity, ECC-protected DDR5 options and QSFP112 networking, with Ethernet and InfiniBand modes. Before evaluating a particular system, check its BlueField-3 part number: memory, port count, connector, clock details and other capabilities vary by SKU.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Which infrastructure workloads it targets
- Networking and virtualization: RDMA, RoCE, SR-IOV, VXLAN, VirtIO, programmable packet processing, cloud overlays and software-defined networking.
- Storage: NVMe over Fabrics (NVMe-oF), RAID and GPUDirect Storage.
- Security: IPsec, TLS, AES-GCM, public-key and regular-expression acceleration, firewalling and tenant isolation.
- HPC and AI communications: GPUDirect RDMA and high-speed Ethernet or InfiniBand traffic processing.
- Other infrastructure: video-streaming processing, time-sensitive communications and telco or edge services.
The list illustrates the DPU’s intended role: accelerating and isolating infrastructure services, not acting as a general-purpose host CPU or a GPU. The exact software and hardware path depends on the application and the chosen BlueField-3 configuration.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the performance claims do—and do not—mean
NVIDIA has described BlueField-3 as capable of offloading data-center services equivalent to as many as 300 CPU cores, and has cited “10x accelerated compute” and “4x cryptography acceleration.” These are NVIDIA claims, not independent comparative benchmark results in the cited Hot Chips material. They should not be read as guaranteed reductions in server core count: realized offload depends on workload, configuration, software and how much of the infrastructure stack is actually moved to the DPU.
NVIDIA’s 2021 preview also said the relevant workloads were growing faster than Moore’s law and already consumed a third of server CPU cycles. That is NVIDIA’s characterization, not a universal measurement for every server fleet. The same preview quoted a Dell’Oro Group forecast that smart networking ports would rise from 4.4 million in 2020 to 7.4 million by 2025. This was a market forecast quoted by NVIDIA in 2021, not a reported current port count.
How to assess whether BlueField-3 fits a system
A useful evaluation starts with the infrastructure function to offload rather than the headline bandwidth number. Compare the target card and deployment against these requirements:
- Network protocol and speed: Determine whether the environment needs Ethernet or InfiniBand and confirm the exact port configuration.
- Host topology: Check the server’s PCIe generation, available lanes, switch topology and physical compatibility against the card.
- Storage and memory: Confirm the required NVMe-oF or RAID path and verify the card’s onboard memory capacity and ECC options.
- Security and isolation: Identify the cryptographic, firewalling and tenant-isolation functions that will run on the DPU, and validate the required software support.
- Software readiness: Check that the intended application or platform supports DOCA and the necessary BlueField-3 features; hardware capability alone does not establish application compatibility.
- Operations: Validate firmware, cooling, power and form-factor requirements for the exact SKU and server.
BlueField-3’s architecture establishes a broad offload platform, but the cited material does not provide a controlled independent benchmark against BlueField-2, other SmartNICs or competing DPUs. Those comparisons require workload-specific measurements and verification of the precise configurations being compared.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
