Nvidia powers Vera Rubin's Extreme Co-Design, evolving from Grace Blackwell Oberon

Nvidia’s Rubin platform expands the “extreme co‑design” concept introduced with the Grace‑Blackwell Oberon rack by integrating six new silicon components into a single, cable‑free compute tray. According to the Nvidia‑sponsored Newsletter, the tray now houses the Rubin GPU, the Vera CPU, an NVLink 6 Switch, ConnectX‑9 NIC, BlueField‑4 DPU, and Spectrum‑6 Ethernet switch, all mounted on a modular “Strata” module that also includes SOC‑AMM memory and a liquid‑cooling subsystem (Newsletter). The NVLink 6 Switch, described as a state‑of‑the‑art scale‑up fabric, replaces the original NVSwitch used in Oberon and promises higher bandwidth per link while eliminating the need for external cabling, a claim echoed by ZDNet’s coverage of the launch (ZDNet). By consolidating these components, Nvidia claims the Rubin rack can operate as a single distributed accelerator, reducing latency between compute, networking, and storage layers.

From a mechanical standpoint, the Rubin design abandons the traditional cable‑laden rack architecture in favor of a “seamless cableless compute tray” that integrates power delivery, cooling manifolds, and high‑density PCB routing within the tray itself (Newsletter). This shift reduces the number of connectors and cable harnesses, which the Newsletter argues improves both thermal performance and assembly efficiency. The liquid‑cooling loop now incorporates cold plates directly on the Rubin GPU and Vera CPU, with quartz‑dot (QD) heat exchangers and manifold routing optimized for the higher thermal envelope of the new silicon (Newsletter). BlueField‑4 and ConnectX‑9 are positioned on the same PCB to share the cooling infrastructure, further simplifying the thermal design.

On the networking front, the Rubin platform introduces a two‑tier fabric: the NVLink 6 backplane provides a high‑speed, low‑latency scale‑up network within each rack, while the Spectrum‑6 Ethernet switch delivers a scale‑out backbone for inter‑rack communication (Newsletter). The NVLink 6 backplane, built on a new silicon switch architecture, supports up to 72 TB/s of aggregate bandwidth, a substantial increase over the 48 TB/s offered by the previous Oberon generation (Newsletter). Spectrum‑6, described as “one of the best Ethernet networking switches” in Nvidia’s portfolio, adds 400 Gb/s ports to enable rapid data movement across hyperscaler clusters (Newsletter). ZDNet notes that this combination could accelerate large language model (LLM) inference and training workloads by reducing data shuffling overhead (ZDNet).

Cost and supply‑chain implications are also front‑and‑center in Nvidia’s rollout. The Newsletter’s newly released VR NVL72 Component Bill of Materials and Power Budget Model estimates a total power draw of roughly 1.2 MW per full rack, with a projected TCO advantage derived from the integrated design and reduced cabling (Newsletter). By controlling the entire silicon stack—from the Vera CPU to the Spectrum‑6 switch—Nvidia positions itself as the sole vendor offering a “best‑in‑class” suite of components, a claim the Newsletter backs with a comparative analysis that finds no rival providing an equivalent combination of GPU, CPU, NIC, DPU, and Ethernet switch (Newsletter). Competitors such as AMD’s MI450X Helios racks, Google’s TPU‑v4, and AWS’s Trainium 3 are mentioned as “catching up” but still lacking the unified silicon ecosystem that Nvidia now touts (Newsletter).

Finally, the Rubin announcement hints at future strategic moves beyond hardware. The Newsletter alludes to “insight into Nvidia’s plans for their Groq IP” and outlines challenges in scaling HBM production for Micron, SK Hynix, and Samsung, suggesting that supply‑chain bottlenecks could shape the market dynamics of the projected $500 billion Rubin buildout (Newsletter). While Nvidia’s GTC 2025 keynote highlighted a 40× performance leap for AI workloads (VentureBeat), the Rubin platform’s holistic integration aims to translate that raw performance into real‑world efficiency gains for hyperscalers and enterprise AI clusters alike.