[News] NVIDIA Unveils NVLink Fusion, Amplifying ASIC Push with MediaTek & Marvell Onboard

At his pre-Computex keynote, Jensen Huang signaled NVIDIA’s expansion beyond GPUs and deeper into the ASIC world. The company unveiled NVLink Fusion on Monday, a new version of its chip-to-chip communication tech now for sale to rival designers. According to its press release, Reuters and Tom’s Hardware, Marvell and MediaTek are already on board to adopt the technology for their custom chip projects.

As per Reuters, NVIDIA’s NVLink, first developed years ago, is the key ingredient behind its high-speed data transfers between chips—like in the GB200, which fuses two Blackwell GPUs with a Grace CPU to handle massive AI workloads.

According to Tom’s Hardware, NVLink is a proprietary interconnect designed for high-speed, direct communication between GPUs and CPUs. The tech delivers up to 14 times the bandwidth of standard PCIe connections, while still utilizing the familiar PCIe electrical interface, the report adds.

Qualcomm, Fujitsu and Big CSPs Poised to Benefit

Tom’s Hardware notes that NVLink was once locked to NVIDIA’s own chips—except for rare cases like IBM. However, that changed in 2022 with NVIDIA NVLink-C2C, and now NVLink Fusion takes it further, letting firms like Fujitsu and Qualcomm plug the high-speed tech into their own CPUs for more flexible AI system designs, the report adds.

NVIDIA says early adopters like MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence are jumping in—using NVLink Fusion to scale up custom silicon for heavy AI training and inference.

It is worth noting that for cloud giants, NVLink Fusion provides a faster route to scale AI factories to millions of GPUs—using any ASIC, NVIDIA’s rack systems, and its full-stack networking platform, as per NVIDIA.

Notably, Qualcomm has reportedly diving into the server race with its own custom data center CPU, and teaming up with NVIDIA’s NVLink ecosystem puts it in prime position to ride the surge of AI demand, according to Tom’s Hardware. The chip titan confirmed last week that its data center CPU launch will kick off through a partnership with Saudi Arabia’s state-backed AI cloud project and HUMAIN AI.

Meanwhile, Tom’s Hardware notes that Broadcom, AMD, and Intel are missing from the NVLink Fusion party. Instead, they’ve joined forces in the Ultra Accelerator Link (UALink) consortium, pushing an open, industry-standard interconnect to challenge NVLink, the report says.

Read more

• [News] NVIDIA Ditches Hopper, Reportedly Turns to Blackwell for China-Specific AI Chip
  
• [News] Four Themes to Watch at Computex 2025 – Beyond Just NVIDIA

(Photo credit: NVIDIA)