Server - HPC _ High Performance Computing - Sub-Project - (2026-03-31)

Why It Matters

Nvidia’s ultra‑dense Kyber blade could redefine AI compute economics, forcing data centers to rethink power, cooling, and hardware standardization strategies.

Key Takeaways

• Nvidia’s Vera Rubin Ultra uses a new Kyber blade architecture
• Blade packs 4 GPU dies, 16 HBM4, up to 1 TB memory
• Vera CPUs feature 8 SOCAMs, 64 LPDDR5X chips, 1.2 TB/s bandwidth
• Power draw per blade estimated at 12‑15 kW, wafer‑scale density
• Mid‑plane switch blade connects 36 compute blades via MVL links

Summary

The meeting focused on Nvidia’s latest high‑performance computing platform – the Vera Rubin Ultra built on the Kyber rack blade architecture. Presenter Andrew walked the team through the hardware specs revealed at GTC, highlighting the shift from the Grace Blackwell B300 design to a denser, vertically‑mounted blade that houses two CPUs, four Vera Rubin Ultra GPUs, and extensive memory subsystems. Key technical details include four reticle‑size GPU dies per blade, each paired with an I/O die, and a jump to 16 HBM4 stacks, effectively doubling memory capacity and bandwidth over the previous generation. The Vera CPUs integrate eight SOCAM modules, each with four LPDDR5X chips, delivering up to 1 TB of DRAM and 1.2 TB/s bandwidth—again roughly twice Grace’s figures. The design also features pluggable memory modules and a sophisticated liquid‑cooling scheme, pushing estimated blade power consumption to 12‑15 kW. During the discussion, participants noted the clever use of a vertical mid‑plane switch blade that links two sets of 18 compute blades via MVL links, eliminating bulky back‑plane cartridges. A brief exchange referenced ARM’s CPU roadmap for inference workloads, underscoring industry interest in CPU‑GPU co‑design for agentic AI. If Nvidia’s density and power metrics hold, a single Kyber rack could approach 680 kW, rivaling wafer‑scale systems like Cerebras. This raises both opportunities for unprecedented AI throughput and challenges around cooling, power provisioning, and ecosystem standardization, prompting the group to consider open‑community standards for future blade designs.

Original Description

Public call recording of Server - HPC _ High Performance Computing sub-project.