The graphics behemoth is letting others piggyback on its precious NVLink tech, a secret weapon behind its AI dominance, by allowing custom rack-scale designs that fuse Nvidia GPUs with third-party CPUs and accelerators. The list of new dance partners is long: Qualcomm, Fujitsu, MediaTek, Marvell, Alchip, and even design tool outfits like Synopsys and Cadence.
This is the first time Nvidia has loosened its grip on NVLink. NVLink was mostly confined to Nvidia’s kit, other than a brief fling with IBM. The new Fusion effort adds a chiplet next to third-party processors to sling data across systems at what Nvidia claims is up to 14 times PCIe’s bandwidth, using the same old electrical interface.
Fujitsu CTO Vivek Mahajan was happy: “Fujitsu’s next-generation processor, FUJITSU-MONAKA, is a 2-nanometer, Arm-based CPU aiming to achieve extreme power efficiency. Directly connecting our technologies to NVIDIA’s architecture marks a monumental step forward in our vision to drive the evolution of AI through world-leading computing technology and paves the way for a new class of scalable, sovereign and sustainable AI systems.”
Qualcomm boss Cristiano Amon said: “With the ability to connect our custom processors to NVIDIA’s rack-scale architecture, we’re advancing our vision of high-performance, energy-efficient computing to the data center.”
Fusion now supports accelerators beyond Nvidia’s own silicon pets, with ASICs and other oddities from Marvell, MediaTek, and Alchip getting the nod. Astera Labs is pitching in with interconnect magic, while the likes of Cadence and Synopsys are chucking in their IP and design tools.
CEO Jensen Huang opened with the kind of line that could power a dozen headlines: “A tectonic shift is underway. For the first time in decades, data centers must be fundamentally rearchitected with AI being fused into every computing platform.”
Nvidia has thrown in new Mission Control software to tie it all together, managing validation and workloads, promising to get customers to market faster.
But not everyone’s in the party tent. Chipzilla, AMD and Broadcom were left outside as their open-source Ultra Accelerator Link (UALink) consortium stares in at the Fusion cabal. UALink hopes to build an open alternative to Nvidia’s tightly held stack.