Bandwidth， Latency， and scalability are most critical metrics for these connectivity protocols Level Scale-Up Scale-Out CPU-Centric Type NVLink (NVIDIA) Open source Cloud Self-Developed NVIDIA Open source Open source NVIDIA Ethernet Metric/Protocols NVLink 5.0 NVLink 6.0 UALink1.0 UALoE SUE Unified Bus 1.0 Unified Bus 2.0 ICI v7(3D Torus)l ICI (TPU 8t， 3D torus) ICI (TPU 8i， Boardfly) NeuronLink v4 InfiniBand XDR PCle6.0 NVLink C2C (200G/lane) Bandwidth per lane/ link 50GB/s per link 100GB/s per link 25GB/s per lane 50GB/s per link 25GB/s per lane 14GB/s 14GB/s 100GB/s per link 200GB/s per link 200GB/s per link 8 GB/s per lane 100GB/s per port 25GB/s per lane 8GB/s per lane 50GB/s per link (one direction) 46 lanes per link 46 lanes per link 16 instances， 4 8 lanes per link， 6 8 lanes per link， 6 links 8 lanes per link， 6 links 128 lanes per xPU Max. Link/ lane per xPU 18 links per GPU 18 links per GPU 0 4 ports， 4 lanes per 36 links per GPU and 1 link per and 2 links per 4 ports 8 ports IX16 18 links per GPU port lanes per instance links per TPU per TPU per TPU (active) NPU NPU C Bandwidth per xPU (full- 1，800 GB/s 3，600 GB/s 800GB/s 3，600 GB/s 3，200GB/s 1，280GB/s 2，560GB/s 1，200 GB/s 2，400 GB/s 2，400 GB/s 2，048GB/s 800GB/s 400GB/s 256GB/s 1800GB/s duplex) bidirectional bidirectional Ibidirectional bidirectional bidirectional bidirectional bidirectiona al bidirectional bidirectional bidirectiona bidirectional bidirectional bidirectional Ibidirectional bidirectional Latency (end-to-end，one 1000-3000ns <200ns <100ns <500ns N/A <400ns <1，000ns 200ns N/A N/A N/A N/A 500-1000ns 100-200ns <100ns (E) way) (RoCEv2) 144 xPUs per 72 GPUs per 1，024 xPUs per 72-256 GPUs per Current scalability 72 GPUs per pod 1024 xPUs 384 NPUs 8，192 NPUs 9，216 TPUs 9，600 TPUs 1，024 TPUs UltraServer(2 100k+ GPUs 100k+ xPUs 164 xPUs 72 GPUs per pod pod Ipod pod racks) <1m with DAC <1m with DAC <1m with DAC <1m with DAC <1m with DAC <3m with DAC <1m with DAC 3-5m with AEC 3-5m with AEC 3-5m with AEC 3-5m with AEC 3-5m with AEC <1m with PCB 130-40cm with PCle 3-5m with ACC 3-5m with AEC <1m with DAC <1m with DAC <1m with DAC <1m with DAC <1m with DAC Intra-Rack Intra-Rack Intra-Cube Intra-Cube Intra-Group Intra-Rack Iretimers Up to 50m with AOC Up to 40km with Interposer to PCB level Signal travel length 3-5m with AEC 3-5m with AEC 3-5m with AEC 3-5m with AEC 3-5m with AEC Within the Board 100m-2km with optics 3-5m with AEC Within the Board Intra-Rack Intra-Rack Intra-Rack Intra-Rack Intra-Rack Up to 200m with Up to 200m with! Several hundred Several hundred m with Several hundred m with 3-5m with AEC optics Intra/Inter- optics S optics m with optics optics S optics Inter-Rack Intra-rack eo Inter-Rack datacenter e Inter-Rack Inter-Rack Inter-Rack Inter-Rack Inter-Rack 2024(GB200) 4Q26 2026 Timeline 2025(GB300) (Vera Rubin) I2H26-2H27 3Q26 2025 4Q26 2025 2Q26 2Q26 2025 2025 2026 1H25 2026 (Tomahawk Ultra) Huawei Google Broadcom， Developers & key vendors NVIDIA AMD，ALAB AMD，Broadcom Broadcom Amazon，AWS NVIDIA ALAB，Amazon NVIDIA Marvell Proprietary Proprietary Proprietary Ecosystem support Open source Open source Open source Proprietary Proprietary Proprietary Proprietary Proprietary Proprietary Proprietary Open source Open source Open by NVLink Fusion Open by NVLink Fusion Open by NVLink Fusion Source： companies reports， public information， Bernstein analysis， Dolphin Research

英偉達在 Exemplar Cloud 中發現 AI 訓練的瓶頸，指出由於配置問題導致 8-12% 的性能差距

推薦閱讀