UALink’s main goal for now’s to supply a substitute for Nvdia’s high-bandwidth, low-latency, direct interconnect expertise for CPU, GPU-to-GPU connectivity, NVLink. NVLink is primarily utilized in InfiniBand-based networks.
Given the spec’s Ethernet heritage, UALink is seen in most circles as working hand-in-hand with the Extremely Ethernet Consortium to assist increase the large Ethernet neighborhood with AI expertise reinforcements. Many members of the UALink group are additionally growing UEC specs that are geared toward growing bodily, hyperlink, transport and software program layer Ethernet advances for the AI connectivity area.
“UALink is on the vanguard of innovation within the synthetic intelligence and machine studying domains, offering an open ecosystem path to a devoted accelerator interconnect leveraging the ever present Ethernet ecosystem,” the UALink group wrote. “By incorporating UALink Switches, accelerators with UALink functionality, can increase the size‐up area, creating extremely‐excessive bandwidth multi‐node accelerator pods. UALink additionally allows a easy software program mannequin by supporting load/retailer operations throughout a complete pod of as much as 1024 accelerators.”
Now a market round UALink must develop to be a real various to Nvidia. Presently, solely UALink member Synopsys has an precise UALink-based product – it rolled out a UALink IP controller, PHY, and verification IP bundle late final yr. Different members are anticipated to observe swimsuit.
“With the discharge of the UALink 200G 1.0 Specification, the UALink Consortium’s member corporations are actively constructing an open ecosystem for scale-up accelerator connectivity,” mentioned Peter Onufryk, UALink Consortium president, in an announcement. “We’re excited to witness the number of options that may quickly be coming into the market and enabling future AI functions.”