Nvidia BlueField-4 with 64 Arm Cores and 800G Networking Announced for 2026 (www.servethehome.com)

NVIDIA announced the BlueField‑4 DPU at GTC DC (available early as part of Vera Rubin platforms in 2026), packing a 64‑core Arm “Grace” CPU, ConnectX‑9 800G networking, and an estimated 126 billion transistors. While full specs aren’t yet public, the company signaled PCIe Gen6 capability and high-bandwidth networking to match modern data‑center needs. Jensen Huang said the DPU will specifically accelerate KV‑cache functions used with large language models (LLMs), complementing Rubin CPX’s prefill work by speeding lookup and retrieval of older conversational context. For the AI/ML community this matters because a beefy DPU with its own many‑core Arm CPU lets operators offload and accelerate control‑plane, networking and LLM inference‑adjacent services (like KV caches, telemetry, security, and data movement) directly on the NIC. That reduces host CPU overhead, lowers latency for cache hits, and can scale AI “factories” more efficiently—important given how much CapEx is tied to NVIDIA‑based clusters. It’s not the first 64‑core Arm DPU, but NVIDIA’s entry, coupled with 800G/ConnectX‑9 and a potential PCIe Gen6 interface, will push adoption across hyperscalers and networking stacks. A related NVIDIA + Nokia AI‑native 6G compute announcement underscores growing interest in pairing high‑performance DPUs with next‑gen telecom infrastructure.