This week at Computex, Nvidia disclosed NVLink Fusion, a program enabling other companies to integrate the high-speed NVLink and NVLink-C2C interfaces. Licensees can integrate an NVLink-C2C block (IP) into their chips or employ an NVLink chiplet to add the interface to their designs. Like other interfaces, it comprises a physical layer (L1) and a control layer (L2).
What is NVLink?
- NVLink is Nvidia’s local interconnect and is available in two versions. NVLink-C2C supports memory coherency and connects two chips. NVLink (non-C2C) can address many peers. For example, Nvidia’s NVL72 system connects 72 GPUs with NVLink by using the company’s NVLink switches. Supporting many chips within a single domain to form a large node, NVLink is a scale-up technology.
- Scaling out (connecting nodes), by contrast, is achieved by the distinct InfiniBand and Ethernet technologies.
- NVLink 5, the most recent version, employs 224 Gbps serdes, grouping them in two pairs to form 800 Gbps ports (400 Gbps in each direction). Few companies offer serdes this fast, and a complete L1 plus L2 solution is required to enable licenses to implement the interface. The most recent NVLink-C2C employs slower serdes.
- UALink is the open-standard NVLink alternative. In theory, it’s just as fast as NVLink. In practice, the 1.0 UALink specification arrived only a month ago, whereas NVLink 5 is in production. AMD will likely employ it in a future Instinct GPU, such as the MI350X, but the availability of UALink switches is unclear.
- Announced NVLink Fusion licensees include companies offering ASIC-design services: Alchip Technologies (that’s Al as in you can call me, not AI as in artificial intelligence), Cadence, Marvell, MediaTek, and Synopsys. Nvidia indicates these companies will create custom AI accelerators. Cadence and Synopsys may also offer relevant EDA tools. Other licensees include Astera (which makes interconnect chips), Fujitsu, and Qualcomm. The latter two intend to integrate NVLink-C2C to connect server CPUs to Nvidia GPUs.
Qualcomm Goes After Servers Again
Qualcomm’s server-processor plans, embedded within the NVLink Fusion announcement, are a bigger story. The phone-chip maker explored this market in the 2010s, disclosing in 2015 a 24-core design based on a custom Arm CPU. Gaining a new high-performance microarchitecture (Oryon) by acquiring Nuvia, Qualcomm is again able to address servers. Merchant-market Arm-based server processors have consistently fizzled (Nvidia Grace is a special case), but proprietary ones such as Amazon’s Graviton and Google’s Axion have taken off. We expect Qualcomm will focus on semicustom chips, pitting its CPU against Arm’s, employing NVLink Fusion to connect to GPUs, and tightening Nvidia’s grasp on AI acceleration.
Bottom Line
NVLink has been vital to Nvidia’s GPU success. It’s unusual for a company to license differentiating technology. We infer that NVLink licensees must employ an Nvidia chip, such as the Blackwell GPU or Grace CPU, at one end of a link. This would ensure Nvidia gets a piece of the action regardless of a licensee’s intent.
We also infer that hyperscalers have requested the ability to mix and match proprietary and Nvidia chips. We foresee pairing a custom CPU with an Nvidia GPU to be more common than the reverse. The CPU can integrate a hyperscaler’s proprietary networking and security technologies. Persuading Nvidia to adopt UALink is a nonstarter, as furtherance of that standard would wedge open the door to AMD and other competing GPU/NPU suppliers.