What is NVLink?
NVLink is NVIDIA's high-speed interconnect technology designed for efficient data transfer between GPUs, CPUs, and other processors in accelerated computing systems. It replaces traditional PCIe connections with significantly higher bandwidth and lower latency, enabling seamless collaboration in high-performance computing (HPC) and AI workloads.
How fast is NVLink compared to PCIe?
Fourth-gen NVLink delivers 900 GB/s bidirectional bandwidth per GPU, 7× faster than PCIe 5.0 (128 GB/s) .
It is 5× more energy-efficient, consuming only 1.3 picojoules per byte transferred.
What are NVLink’s key features?
Direct GPU-to-GPU/CPU communication: Bypasses PCIe bottlenecks.
Hardware cache coherence: Ensures data consistency between CPUs and GPUs (from NVLink 2.0 onward).
Scalability: Supports large clusters via NVSwitch (e.g., 256 GPUs at 1.8 TB/s).
Where is NVLink used?
AI/Deep Learning: Accelerates multi-GPU model training (e.g., trillion-parameter AI models).
Supercomputers: Power systems like Summit and Sierra.
Data Centers: Optimizes cloud workloads (e.g., NVIDIA DGX servers).
Does NVLink support consumer GPUs?
No. After the RTX 30-series, Ada Lovelace architecture (e.g., RTX 4090) dropped NVLink support to prioritize AI-focused silicon 49. It remains exclusive to enterprise GPUs (e.g., H100, Grace Hopper).
How does NVLink evolve?
| Generation | Bandwidth | Key Innovations |
|---|---|---|
| NVLink 1.0 | 160 GB/s | GPU-GPU links (Pascal) |
| NVLink 2.0 | 300 GB/s | CPU-GPU cache coherence (Volta) |
| NVLink 4.0 | 900 GB/s | NVSwitch integration, C2C packaging |
Why does NVLink matter?
It transforms multi-GPU systems into "one giant GPU," eliminating communication bottlenecks for HPC and AI.
Tag：NVLink
