Samsung has expressed interest in joining the UALink Alliance, which aims to create new interconnect standards for various accelerators, according to DigiTimes. The report suggests that the plan is designed to boost Samsung's foundry business, allowing it to better address its customers' needs. However, Samsung could also develop its own accelerators and connect them using industry standard technology.
At the Samsung Foundry Forum (SFF) 2024, TaeJoong Song, who leads business development at Samsung's foundry division, indicated that Samsung is intrigued by UALink's efforts to standardize AI chip interconnectivity. He mentioned that Samsung is exploring ways to support and possibly join the alliance.
The UALink initiative aims to establish an open standard for more efficient communication between AI accelerators by facilitating direct data transfers between the memory of attached accelerators, thus improving performance and efficiency for demanding compute workloads. The initial UALink specification, version 1.0, will support connecting up to 1,024 accelerators within an AI computing pod, creating a reliable, scalable, and low-latency network. The technology will, of course, challenge Nvidia's NVLink, which serves the same purpose.
While Samsung develops its own system-on-chips for smartphones and SSDs, it doesn't build processors or accelerators akin to AMD's Instinct MI300A or Instinct MI300X that genuinely require a high-performance, low-latency scalable interconnect. While Samsung may need the UALink technology for its processors at some point, it is more likely that the company's foundry unit needs to be a part of the alliance to make chips with UALink for its customers.
The UALink Alliance consists of industry giants like AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft. All of them are potential customers for Samsung Foundry, so the company certainly needs to access the patents on which UALink relies.
Despite being competitors in other areas, the companies in the UALink Alliance are collaborating to reduce their reliance on Nvidia's NVLink. This collective effort highlights their urgency in addressing Nvidia's monopolistic hold on AI chip interconnect technology.
Samsung's participation in the UALink technology could attract orders for next-generation AI chips from major tech companies. Samsung has already secured orders from AI chip startups such as Tenstorrent, Groq, Rebellions, and DEEPX for its 4nm process at its upcoming fab in Taylor, Texas. Additionally, Samsung Chairman Jae-yong Lee recently met with CEOs of Amazon, Meta, and Qualcomm, hinting at potential AI chip foundry deals.
