Nvidia launches the BlueField-4. This high-throughput DPU is designed to handle network-related workloads, relieving servers.
Nvidia launches BlueField-4. This Data Processing Unit (DPU) succeeds BlueField-3. The throughput speed of the new DPU doubles from 400 Gb/s to 800 Gb/s. The launch at the GTC event in Washington is no surprise: Nvidia is executing its existing roadmap neatly and already spoke about the chip in 2021.
Specifications
The Nvidia BlueField-4 combines an ARM-based Grace CPU with ConnectX-9 networking capabilities. The CPU has 64 Arm Neoverse V2 cores and the chip comes with 128 GB LPDDR5.
Connection to the host is via the PCIe Gen6 interface. The chip is intended for servers with RTX Pro GPUs as well as powerful GB200- and GB300-based systems and Nvidia DGX clusters.
CPU, GPU, and DPU
Nvidia developed the DPU with Spectrum-X Ethernet in mind. This technology works somewhat like an HPC interconnect, linking servers together into larger clusters. The InfiniBand interconnect is also supported.
Nvidia sees BlueField as an essential building block for its so-called ‘AI factories’: HPC clusters tailored for AI workloads. These clusters combine CPUs for general management with GPU accelerators like the Blackwell chips and DPUs like BlueField-4.