Arm brings Nvidia’s NVLink tech to its AI data center chips

Arm and Nvidia have expanded their partnership to enable Arm-based compute integration into Nvidia’s NVLink Fusion ecosystem, the companies announced.

This collaboration allows ecosystem partners to connect Arm’s Neoverse platform with Nvidia’s NVLink Fusion, aiming for higher bandwidth and full coherency between CPUs, GPUs, and accelerators in AI data centers.

Arm, a UK-based semiconductor and software design company, said its Neoverse platform now supports the latest AMBA CHI C2C protocol to ensure compatibility with NVLink Fusion, allowing for more flexible and efficient AI system designs.

Nvidia’s NVLink Fusion is designed to address memory and bandwidth bottlenecks in AI infrastructure by creating a unified rack-scale architecture.

Both companies said the move would give data center partners more options in connecting workload accelerators to Arm-based systems.

🔗 Source: Arm

🧠 Food for thought

Implications, context, and why it matters.

Which Neoverse cores and OEMs will actually ship NVLink Fusion systems

Adoption timelines remain missing 1. Buyers do not know which original equipment manufacturer (OEM) partners will ship systems, which Arm Neoverse (Arm’s data center CPU core family) designs like V2 or N2 will support NVIDIA’s NVLink Fusion interconnect, or when hardware will be for sale.
AWS Graviton4, Google Axion, and Microsoft Azure Cobalt 100 use Neoverse cores 2. No public NVLink Fusion integration timelines exist for these chips 1.
Fujitsu’s 144-core Monaka CPU from the Japanese computing vendor targets NVLink Fusion integration 3. No ship dates or Neoverse platform details are public, so the plan could be near term or a multi-year effort that may not ship.
Separate Arm’s protocol specification from actual silicon tape-outs, when a chip design is finalized and sent to manufacturing, and from system availability 4. Protocol support does not ensure partners will ship products given the complexity of NVLink chip-to-chip (C2C) die-to-die or chiplet interfaces linking multiple silicon dies within a package.

Cooling and interconnect implications of NVLink-based rack-scale systems

NVLink with the NVLink 5 Switch can link up to 576 fully connected GPUs in a non-blocking fabric, where each device can communicate at full bandwidth 5. The switch has 144 ports and 14.4 TB/s capacity.
Vendors of liquid cooling and high-speed copper or optical cabling may see rising demand as hyperscalers (large cloud providers) plan NVLink-based racks with higher power use 6. Custom backplanes (chassis-level interconnect boards) and thermal interface materials (compounds that improve heat transfer between chips and coolers) could benefit too. NVIDIA sells full rack solutions, yet OEM ecosystems still supply specialized parts.
Validation tool providers that build electronic test equipment and verification software can tackle new testing needs for custom Arm CPUs with NVLink-C2C connectivity 4. Ensuring coherency, bandwidth, and reliability at rack scale (spanning an entire server rack) may require gear beyond typical server validation.
Cooling firms need clear power and thermal targets for each NVLink-based setup to size offerings well 5. That detail also helps time outreach to data center operators planning Arm-based AI infrastructure.