NVIDIA CX7 BlueField-3 DPU with GPU Offload (ConnectX-7)

The NVIDIA BlueField-3 DPU (Data Processing Unit) with ConnectX-7 networking fabric represents a converged platform that combines a high-performance SmartNIC with an integrated Arm-based compute complex, purpose-built for offloading, accelerating, and isolating infrastructure workloads from the host CPU. BlueField-3 integrates the ConnectX-7 network controller — supporting PCIe Gen 5.0 and delivering up to 400GbE or NDR InfiniBand connectivity — with an array of Arm Cortex-A78 cores, hardware-accelerated cryptographic and compression engines, and dedicated RegEx processing, allowing storage, networking, and security functions to execute entirely on the DPU without consuming host CPU cycles.

The GPU Offload capability (referenced in the product designation) reflects BlueField-3's position within NVIDIA's accelerated computing ecosystem, enabling direct DMA data paths between the DPU's network interface and connected NVIDIA GPUs. This GPUDirect RDMA architecture is critical for AI and HPC workloads where data must move from the network fabric directly into GPU memory at line rate, bypassing the CPU and system memory to dramatically reduce latency and eliminate bottlenecks in distributed training and inference pipelines. The ConnectX-7 fabric controller underpins this with support for RDMA over Converged Ethernet (RoCE) and InfiniBand transport protocols.

Manufactured under Dell's program with NVIDIA and carrying manufacturer part number 900-9D3B6-0007-ST0, this DPU targets modern enterprise data centers deploying disaggregated, software-defined infrastructure. It is particularly well-suited for environments running VMware vSphere with VCF, Red Hat OpenShift, or bare-metal Kubernetes, where tenant isolation, zero-trust microsegmentation, and storage acceleration (NVMe-oF, vStorage APIs) are architectural requirements. Omnixon Global makes this unit available to enterprise buyers across UAE, GCC, EMEA, and APAC regions.

Ideal for

• AI and HPC distributed training acceleration via GPUDirect RDMA, enabling direct high-bandwidth, low-latency data transfers between the network fabric and GPU memory without CPU involvement
• Software-defined networking and security policy enforcement offloaded entirely to the DPU, freeing host CPU cores for application workloads while achieving consistent microsegmentation and zero-trust isolation
• NVMe over Fabrics (NVMe-oF) storage target and initiator offload, enabling disaggregated storage architectures with line-rate throughput and reduced host CPU overhead
• Telco and 5G vRAN infrastructure, leveraging BlueField-3's hardware packet processing, timing, and acceleration engines to meet real-time Open RAN performance requirements
• Secure multi-tenant cloud hosting, where the DPU enforces tenant network and storage isolation at the hardware level, preventing cross-tenant data access independent of the host hypervisor
• Data center infrastructure management (DOCA-based), enabling out-of-band lifecycle management, intrusion detection, and encrypted traffic inspection running isolated from the host OS

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.