Nvidia Delays Groundbreaking Kyber Rack To 2028 Amid Manufacturing Hurdles

Nvidia’s Kyber Rack Delayed to 2028: A Delicate Balance Between Innovation and Manufacturing Challenges

According to recent reports from analyst firm SemiAnalysis, Nvidia’s highly anticipated Kyber rack for its Rubin Ultra GPUs has been delayed to 2028. This delay is attributed to manufacturing challenges with the PCB (Printed Circuit Board) midplane, which connects eight Oberon racks between the NVSwitches, a crucial component of the system.

The cancellation of the stopgap solution, NVL72x2, has left Nvidia without a proven alternative to expand the scale-up capabilities of Rubin Ultra in 2027. To understand the significance of this delay, it’s essential to delve into the technical details behind Kyber and its orthogonally mounted compute trays.

The system is designed to stack 144 Rubin Ultra packages, doubling the 72 packages found in a current Oberon NVL72 rack. This configuration enables every GPU-to-GPU link inside the cabinet to run through a rigid board carrying the all-copper NVLink fabric. The use of liquid cooling by default and the stacking of GPUs contribute to the system’s impressive density.

However, this innovative design comes with its own set of challenges. The orthogonal backplane, which sits between Kyber’s vertically mounted compute trays and the switch trays behind them, replaces cable harnesses with a rigid board that carries the NVLink fabric. This approach offers several benefits, including reduced signal loss due to copper traces, improved power delivery, and enhanced thermal design.

The manufacturing of this critical component has proven to be a significant hurdle. Trade analyses reveal three 26-layer sections laminated into one 78-layer stack, close to a square meter in area, with trace spacing at or below 25μm and impedance held within a tolerance of 5% to maintain 448 Gb/s-class signaling intact. A cabled version of the same interconnect would require upward of 20,000 discrete cables, highlighting the advantages of moving the wiring onto a single passive board.

Despite these technical challenges, Nvidia had envisioned NVL72x2 as a stopgap solution to widen Rubin Ultra’s scale-up in 2027. However, after its largest customers expressed concerns about running two linked cabinets as a single unit, the company has abandoned this plan. Instead, NVL576, a separate configuration tying eight racks together through co-packaged optics, is likely to slip into production or ship in low volume until that optical technology matures.

The cancellation of NVL72x2 leaves Nvidia with “no proven solution” to expand the scale-up world size for Rubin Ultra, limiting the largest single Rubin Ultra domain in 2027 to match but not exceed what Oberon already delivers. This development underscores the complexity and risks associated with developing cutting-edge technologies like Kyber.

In a related move, Nvidia has dropped the quad-chiplet Rubin Ultra GPU for a dual-chiplet part due to manufacturing execution concerns, halving the accelerator’s per-package compute. Furthermore, SemiAnalysis predicts that a fully production-ready co-packaged optics NVSwitch will not be available until the Feynman generation, which follows Rubin. This leaves copper as the only near-term solution for linking Rubin Ultra at rack scale, further emphasizing the importance of addressing the PCB midplane challenges.

Nvidia’s decision to delay Kyber and cancel NVL72x2 is a sobering reminder that innovation often requires compromise between technological advancements and manufacturing feasibility. As the company continues to push the boundaries of GPU design and manufacturing, it is essential to acknowledge these trade-offs and strive for more efficient solutions in the future.

The delay of Kyber has sparked debate among industry experts, with some questioning whether Nvidia’s aggressive expansion plans are unsustainable. However, others argue that the company’s commitment to innovation and its reputation for delivering high-performance products make the delay a minor setback.

Ultimately, the success of Kyber and Rubin Ultra will depend on Nvidia’s ability to navigate these technical challenges and overcome manufacturing hurdles. As the industry continues to evolve, it is crucial to monitor the development of this technology and assess the implications of such delays on the broader market.

Nvidia’s commitment to pushing the boundaries of GPU design and manufacturing will undoubtedly have significant consequences for the industry. The company’s ability to adapt to these challenges and find innovative solutions will be key to its continued success.

Original Source

• Read the full article here