Intel’s last lifeline: Gaudi 3 meets NVIDIA’s Blackwell in a hybrid rack-scale construct

Intel’s desperate attempt to gain a foothold in the booming AI market has taken on a new, almost paradoxical form, report our colleagues at Wccftech: instead of remaining on a confrontational course with industry leader NVIDIA, the company is now integrating itself into its ecosystem, of all things. The new rack-scale system based around the company’s own Gaudi 3 accelerators combines Intel’s server hardware with NVIDIA’s Blackwell GPU and the ConnectX network stack. It is a move that reveals as much about Intel’s current weakness as it does about NVIDIA’s dominant position.

In concrete terms, the system looks like this: Each compute tray contains two Xeon processors, four Gaudi 3 chips, four NVIDIA ConnectX-7 NICs (400 GbE) and a BlueField-3 DPU. These trays are connected rack by rack to Broadcom’s Tomahawk 5 switches, allowing up to 51.2 Tb/s of throughput. A total of 16 trays per rack provide the backbone of this platform. Blackwell handles the complex prefill phase, i.e. the massive, contextual matrix processing, while Gaudi is responsible for the latency-critical decode stage. The division of labor follows technical common sense. Blackwell is optimized for raw computing power and memory bandwidth, while Gaudi, with its Ethernet architecture, favours scalable communication across many nodes. Intel promises up to 1.7 times the acceleration of a pure B200 system in this hybrid setup, but only for small, dense models and so far without independent benchmarks.

What sounds like innovation is, on closer inspection, a strategic retreat. The Gaudi platform was never a commercial hit. Too inconsistent, too late on the market, with a software stack that is always one or two generations behind. The fact that even part of the system now relies on NVIDIA software and infrastructure is an admission: Intel can no longer serve the market alone. Even within its own rack, it is handing over key tasks to its direct competitor. This is a double victory for NVIDIA. On the one hand, it sells its own GPUs and network components, and on the other, it cements its own platform as an indispensable standard. Anyone who does not live entirely in the NVIDIA universe will have to adopt it, at least in part, and thus Intel’s actual competitive advantage, namely independence and open standards, is undermined.

It remains to be seen how far this hybrid rack solution really scales. Without a competitive software ecosystem, Gaudi will remain a supplier, not the pacesetter. And since Intel plans to phase out the architecture in the foreseeable future anyway, it is questionable how sustainable the current solution really is. Rather, it looks like a last attempt to monetize remaining stocks before the next generation, perhaps with Falcon Shores, takes over.

Intel will sell its product in an NVIDIA package in the future, a technological horse-trading deal that could not be clearer about the balance of power in the AI sector.

Source: Wccftech