CoCoLink 推出 PCIe 伺服器,作為 Nvidia NVLink 的替代方案
CoCoLink Unveils PCIe Server as Alternative to Nvidia NVLink
CoCoLink 推出採用 Multipath PCIe Fabric(MPPF)的 AI 伺服器,以 PCIe 連接多個加速器,無需依賴 Nvidia 專有的 NVLink。旗艦系統 Klimax-6155 可在單部 6U 伺服器容納 15 個加速器,每個加速器的數據傳輸頻寬為 52 至 53 GB/s。公司亦公佈 ePIM-MeDUSA 路線圖,計劃於 2029 年提供樣品、2030 年量產。

CoCoLink has unveiled a server that connects multiple accelerators without using Nvidia’s proprietary NVLink interconnect technology.
CoCoLink participated in SC26, an international supercomputing conference and exhibition held in Chicago from November 15 to 20, local time, showcasing its PCIe-based artificial intelligence (AI) servers and development roadmap for next-generation accelerators.
CoCoLink’s technology, known as Multipath PCIe Fabric (MPPF), uses PCIe, a high-speed data communication standard for connecting internal computer components. It employs multiple PCIe switches to increase data transmission paths, enabling connections between numerous accelerators without relying on Nvidia’s proprietary NVLink technology.
Its flagship Klimax-6155 system accommodates 15 accelerators in a single 6U server. Equipped with a 256-lane PCIe 5.0 switch, it provides data transfer bandwidth of 52 to 53 gigabytes per second (GB/s) per accelerator. The system has a power consumption rating of 16 kilowatts (kW) and supports direct liquid cooling (DLC).
Another model, the Klimax-4082, houses eight accelerators in a 4U server. It uses a 144-lane switch and has a power consumption rating of 8 kW. The system was listed in MLPerf Training v6.0, an AI training performance benchmark released in June.
CoCoLink also unveiled its development roadmap for the ePIM-MeDUSA, a next-generation computing accelerator developed in-house. The company is targeting double-precision floating-point (FP64) performance of 236 teraflops and 4-bit floating-point (FP4) performance of 15.1 petaflops for AI workloads.
The accelerator is planned to feature 256 gigabytes (GB) of GDDR7 memory with bandwidth of 4.61 terabytes per second (TB/s). It will support PCIe 6.0 and Compute Express Link (CXL) 3.0, with a target power consumption of 1,400 watts. CoCoLink aims to begin supplying samples in 2029 and start mass production in 2030.
來源:The Elec · thelec.net