Nvidia unveils Vera Rubin
Nvidia announced the Vera Rubin platform — a full‑stack AI infrastructure push that bundles CPUs, GPUs, interconnect, and data processing into unified systems. That formalizes a hardware-software stack approach and raises the integration bar for orchestration and virtualization layers. (networkworld.com)
NVIDIA framed Vera Rubin as “seven new chips in full production” and shipped the platform as five rack configurations — Vera Rubin NVL72 GPU racks, Vera CPU racks, Groq 3 LPX inference racks, BlueField‑4 STX storage racks and Spectrum‑6 SPX Ethernet racks. (nvidianews.nvidia.com) The silicon lineup named in the announcement includes the Rubin GPU, Vera CPU, NVLink™ 6 switch, ConnectX‑9 SuperNIC, BlueField‑4 DPU and Spectrum‑6 Ethernet switch — with the Groq 3 LPU added as an integrated inference accelerator. (nvidianews.nvidia.com) NVIDIA’s Rubin GPU is described in vendor and press briefings as a dual‑die part with 288 GB of HBM4, roughly 22 TB/s of memory bandwidth, and about 50 PFLOPS of NVFP4 inference throughput (≈35 PFLOPS training performance reported). (tomshardware.com) The flagship NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs and is quoted as delivering roughly 3.6 exaFLOPS NVFP4 inference and 2.5 exaFLOPS training at rack scale in NVIDIA’s NVL72 reference architecture. (videocardz.com) NVIDIA named OpenAI, Anthropic, Meta and Mistral among early supporters and said “every major cloud provider” is backing the Vera Rubin roll‑out in partner statements and coverage. (venturebeat.com) In his GTC keynote, Jensen Huang tied Rubin and Vera to the company’s revenue runway, saying demand could bring about $1 trillion in orders for Blackwell and Vera Rubin systems through 2027. (cnbc.com) NVIDIA also emphasized heterogeneous racks for agentic inference — integrating Groq 3 LPUs for low‑latency workloads — and third‑party reporting shows DGX Rubin NVL8 systems will use Intel Xeon 6 as the host CPU in some configurations. (computerweekly.com / crn.com)
