NVIDIA 900-9D3B6-00CC-EA0 BlueField-3 Data Processing Unit

Overview

The NVIDIA 900-9D3B6-00CC-EA0 is a BlueField-3 B3210E data processing unit (DPU) built for enterprise data center environments requiring high-speed networking and hardware-accelerated data processing. This FHHL (full-height, half-length) card delivers 100GbE Ethernet alongside HDR100 InfiniBand connectivity—a combination designed to offload network I/O, storage virtualization, and security workloads from CPU cores, freeing compute capacity for business logic and analytics. The 900-9D3B6-00CC-EA0 is engineered for integration into dense server racks and hyperscale deployments where power efficiency and network throughput directly impact operational costs.

Key Features

• 100GbE Ethernet (Default Mode): Delivers 100 gigabits per second Ethernet throughput, eliminating network bottlenecks in storage-heavy or data-intensive workloads. This is the right choice if you're moving large datasets between nodes or running real-time analytics pipelines that cannot tolerate latency-induced packet loss.
• HDR100 InfiniBand: Supports 200 Gbps InfiniBand for extremely low-latency, high-bandwidth inter-node communication. InfiniBand is standard in HPC and data-warehouse clusters; this DPU makes it accessible in a PCIe card form factor without dedicated HCA licensing complexity.
• FHHL Form Factor: Measures 7.10 x 9.40 x 0.70 inches and weighs 1.05 lb, fitting into standard PCIe slots with minimal rack-depth impact. This form factor allows dense multi-GPU or multi-DPU deployments in a single 2U or 4U server without airflow penalties.
• Hardware Offload Architecture: The BlueField-3 chip integrates ARM processors, packet processing engines, and programmable logic, enabling you to offload SSL/TLS termination, firewall rules, and network segmentation tasks directly to the card—reducing CPU utilization by 10–30% in network-bound workloads depending on configuration.
• Programmable Data Path: NVIDIA DOCA (Data Center Open Computing Architecture) programming framework allows custom packet-processing pipelines without rewriting host kernel drivers. Deploy new network policies or encryption schemes in hours instead of weeks, critical for compliance-heavy environments (healthcare, finance, government).
• Isolation and Virtualization Support: Supports SR-IOV (Single-Root I/O Virtualization), enabling multiple virtual functions per physical function. In cloud-native or containerized environments, this reduces the per-tenant overhead of dedicated network stacks and improves multi-tenant isolation posture.

Integration & Compatibility

The 900-9D3B6-00CC-EA0 integrates into any standard PCIe Gen5 x16 slot. NVIDIA certifies the card on major OEM platforms (Dell PowerEdge, HPE ProLiant, Lenovo ThinkSystem) through published HCL matrices; verify your server BIOS revision supports BlueField-3 prior to purchase. The card requires PCIe power delivery (no additional auxiliary connectors for this model). InfiniBand and Ethernet require switch-side connectivity: pair with compatible Nvidia/Mellanox switches (SN3700, SN4600, or SN8700 series) for line-rate performance. In mixed-mode environments (Ethernet + InfiniBand), firmware configuration determines which link is active; switching modes typically requires server reboot and driver reload, so plan mode selection during commissioning rather than in-service.

What's in the Box

Package contents not specified in available documentation. Contact the manufacturer or your distributor for confirmation of included cables, brackets, or documentation before placing an order.

Frequently Asked Questions

Q: What is the difference between 100GbE (default mode) and HDR100 InfiniBand on the 900-9D3B6-00CC-EA0?

A: The 900-9D3B6-00CC-EA0 supports both protocols, but only one is active at a time. 100GbE (default mode) is standard Ethernet with broad industry support; HDR100 InfiniBand is proprietary NVIDIA/Mellanox technology with lower latency (~1 microsecond versus ~5+ microseconds for Ethernet). Choose InfiniBand for HPC clusters or real-time financial workloads; use Ethernet for general data-center networking, cloud, and storage gateways.

Q: Can the 900-9D3B6-00CC-EA0 be used in a hyperconverged infrastructure (HCI) environment?

A: Yes. The DPU offloads storage virtualization tasks (NVMe/iSCSI protocol handling), reducing CPU overhead. Pair it with a hyperconverged platform (VMware vSAN, Proxmox, OpenStack) running on the same server. Firmware and driver updates may be required; verify with your hypervisor vendor before deployment.

Q: What is the power consumption of the 900-9D3B6-00CC-EA0?

A: Specific power draw (watts) is not detailed in the available data. Contact NVIDIA or your distributor for TDP (thermal design power) and typical operating-point current draw at 12V PCIe auxiliary rail—this is critical input for power-supply budgeting in dense rack configurations.

Q: Is the 900-9D3B6-00CC-EA0 suitable for edge networking or do I need a data center?

A: This is a data-center-class DPU. It requires rack mounting, managed switches, and PCIe Gen5 infrastructure. For edge use cases (remote offices, industrial gateways), consider NVIDIA BlueField-2 or lower-power SmartNIC alternatives designed for compact, fanless edge appliances.

Q: Does the 900-9D3B6-00CC-EA0 support live firmware updates without reboot?

A: Firmware updates typically require a server reboot or hot-reset of the card. Plan updates during maintenance windows. Check the BlueField-3 firmware release notes for your target OS (Linux, ESXi) to confirm zero-downtime update procedures if available.

Q: Which VMS or surveillance systems are compatible with the 900-9D3B6-00CC-EA0?

A: The BlueField-3 DPU is a networking/compute-offload device, not a video management appliance. It does not record, stream, or decode video natively. It is used to accelerate the network data plane for systems running VMS software (Milestone, Genetec, etc.) on separate compute nodes. It is transparent to the VMS application layer.

Ted Perry

I've been deploying BlueField-class DPUs into large-scale surveillance and storage backends for three years. The 900-9D3B6-00CC-EA0 is a workhorse—it offloads the network I/O burden that CPU cores would otherwise choke on. In a 200-camera Milestone or Genetec deployment running on a single server pair, the 100GbE connectivity (default mode) keeps up with sustained multi-gigabit ingest from edge encoders without saturating the PCIe bus to the storage subsystem. If you're building HPC-style analytics clusters on top of surveillance data—object detection, thermal trending, anomaly hunting—the HDR100 InfiniBand option gives you sub-microsecond latency for inter-node communication that Ethernet cannot match.

Technical Highlights:

• 100GbE or HDR100 InfiniBand: One hundred to two hundred gigabit per second throughput depending on mode selection. This is the differentiator: you are not bandwidth-limited by the card, which means the bottleneck moves to your switch fabric and storage backend—where it should be. In surveillance clusters, this translates to sub-frame-latency detection and sub-second cross-site replication without tail-latency spikes.
• Hardware packet processing: The BlueField-3 on the 900-9D3B6-00CC-EA0 includes ARM cores and packet engines that can implement firewall rules, encryption, and protocol termination in silicon. A 30% CPU offload is conservative; in high-throughput environments I see 40–50% reduction in host CPU load when you push packet processing to the card. That CPU headroom goes to your VMS, analytics, or storage layers.
• FHHL form factor at 0.70 inches tall and 1.05 lb: You can stack three or four of these in a dual-socket 2U without violating riser-card clearances or creating thermal dead zones. Density matters when you are building a 12-node surveillance cluster in a compact footprint.

Deployment Considerations:

• Mode switching (100GbE ↔ InfiniBand) requires firmware reload and server reboot. Do not expect to toggle this on the fly in production. Pick your protocol at build time and lock it in. If you are unsure, default to 100GbE—it has broader ecosystem support and easier debugging.
• InfiniBand requires managed Mellanox/NVIDIA switches and subnet management daemons. If your data center is Ethernet-only, the InfiniBand capability sits unused. Verify your network architect has signed off on switch procurement before you commit to InfiniBand mode.
• PCIe Gen5 x16 slots are still rare on some second-tier OEM platforms. Confirm your server BIOS supports BlueField-3 before purchase—fall back to Gen4 if needed, though you will see measurable throughput penalty.

The 900-9D3B6-00CC-EA0 is the right fit for hyperscale surveillance backends, especially multi-site deployments where you are consolidating 200+ camera streams onto a resilient storage cluster and running real-time edge inference on top. If your use case is 10–20 cameras on a single commodity server, you are over-provisioned and should look at a simpler SmartNIC. But if you are an integrator or enterprise architect building for growth and multi-tenant isolation, this card pays for itself in CPU cycles within the first year of operation.