Nvidia drops quad-die Rubin Ultra for dual-die

Tom's Hardware reports that Nvidia canceled a quad-die variant of the Rubin Ultra GPU in favor of a dual-die design, citing "manufacturing execution concerns" in its June 30, 2026 coverage. Tom's Hardware states the quad-die package would have linked four near-reticle-sized compute dies and supported 16 HBM4E stacks. Separate reporting by Wccftech, which cites Taiwanese supply-chain sources, describes the same revision as a move away from a single CoWoS-L-style multi-die package toward a board-level approach where multiple Rubin dies appear across a server blade - a reported 2+2 arrangement on a Kyber blade. Tweaktown corroborates the dual-die revision and attributes it to warping and thermal issues in the original design.

Multi-die CoWoS-L style integration increases engineering friction in several ways: advanced interposer and through-silicon-via routing, concentrated thermal density, and mechanical stress that can cause warping and yield loss. These are industry-wide packaging failure modes and are not unique to the companies in the reports. Board-level assembly spreads thermal and mechanical load across PCB mounts and allows standard manufacturing flows for memory stacks like HBM4, at the cost of denser interconnect latency and cabling complexity.

Reporting frames the revision as supply-chain and manufacturability-driven rather than purely architectural. That pattern - where vendors trade extreme single-package integration for more modular, rack-scale assembly - has precedent in server accelerator design when yields or supplier readiness lag aggressive packaging roadmaps.

Neither Tom's Hardware nor Wccftech cites an Nvidia public statement explaining the rationale, and the two outlets disagree on the performance consequence: Tom's Hardware suggests the scaled-back package could run roughly half as powerful as the original quad-die design, while Wccftech's industry sources expect board-level assembly to preserve aggregate compute and HBM capacity per blade. Teams planning around Rubin Ultra procurement timelines should treat current specs as provisional and confirm die count, HBM capacity, and power/cooling requirements once Nvidia publishes official specifications.

• •Vendor statements or supplier confirmations about CoWoS-L or alternate packaging routes.
• •Official Rubin Ultra specs confirming die count, HBM4E capacity, or Kyber blade design.
• •Indicators of yield or thermal issues from packaging partners reported by Taiwanese supply-chain outlets.

Tom's Hardware, Wccftech, and Tweaktown are the primary sources for these claims, and all three describe the change consistently even though they diverge on the performance implications. Until Nvidia confirms specifications directly, the report should be read as credible industry reporting rather than an official disclosure - the packaging shift itself (single-package to board-level) is corroborated across sources, but the performance-impact question remains open.