https://www.linkedin.com/company/network-switch/
Answer first: NVLink is NVIDIA's proprietary interconnect for supported CPU/GPU and GPU/GPU domains, while NVSwitch builds a switched NVLink fabric; compare only the exact architecture, system topology, link count, direction convention, software, workload, and documented peak conditions. Review NVIDIA's Hopper NVLink/NVSwitch overview and current DGX GB rack-system guide. Continue with InfiniBand cable guide, NIC selection guide, scale-up versus scale-out architecture, RDMA transport and deployment, SmartNIC and DPU selection, InfiniBand architecture and selection. Evidence boundary: architecture names, vendor peak rates, preserved scenarios, and roadmap statements are not independent workload benchmarks or guaranteed application outcomes; latency, throughput, scaling, utilization, availability, power, and cost depend on exact hardware, software, topology, message sizes, collective patterns, congestion, placement, cooling, and test method. Procurement boundary: verify exact compute, GPU, CPU, NIC, HCA, DPU, switch, cable and optics PIDs, firmware, drivers, SDKs, licenses, topology, compatibility, power, cooling, lifecycle, support, stock, delivery, and acceptance tests in writing.
To address this, NVIDIA introduced NVLink, a proprietary high-speed interconnect designed to deliver low latency, high bandwidth, and efficient GPU scaling.
This article explores the evolution of NVLink from its first release in 2014 to its fourth generation, compares it with PCIe, explains the role of NVSwitch, and analyzes its role in AI, HPC, and the future of optical interconnects.
What is NVLink and Why it Matters?
The Bandwidth Bottleneck of PCIe
Peripheral Component Interconnect Express (PCIe) has been the standard bus connecting GPUs and CPUs for decades. While PCIe has advanced from Gen 1 (2.5 Gbps per lane) to Gen 6 (64 Gbps per lane, PAM4), its scaling could not keep pace with the exponential growth of GPU performance.
For instance, PCIe 3.0 x16 delivers ~16 GB/s bandwidth, but modern GPUs can consume data at hundreds of GB/s. This mismatch created a bottleneck in multi-GPU systems, where inter-GPU communication became a limiting factor.
Introduction to NVLink
NVLink is NVIDIA’s proprietary point-to-point interconnect that bypasses PCIe switches and CPU scheduling, allowing GPUs to communicate directly with bidirectional bandwidths up to 900 GB/s (NVLink 4.0).
Key design goals of NVLink:
- High bandwidth: Multiplying effective GPU interconnect capacity.
- Low latency: Shorter data paths compared to PCIe switching.
- Scalability: Enabling efficient GPU meshes in DGX systems.
NVLink Generations and Performance Improvements
NVIDIA has released four generations of NVLink, each doubling (or nearly doubling) bandwidth.
NVLink 1.0 (2014, Pascal P100)
- Implemented on Pascal GPUs (P100).
- 4 links per GPU, each with 8 lanes.
- Lane speed: 20 Gbps.
- Total bidirectional bandwidth: 160 GB/s, ~5× faster than PCIe 3.0 x16.
- Supported DGX-1, with up to 8 GPUs in a cube mesh topology (not fully connected).
NVLink 2.0 (2017, Volta V100)
- Released with Volta GPUs (V100).
- 6 links per GPU, lane speed 25 Gbps.
- Total bidirectional bandwidth: 300 GB/s.
- Introduction of NVSwitch 1.0 (18 ports, 50 GB/s per port).
- Enabled DGX-2, with 16 fully interconnected GPUs.
NVLink 3.0 (2020, Ampere A100)
- Introduced with Ampere GPUs (A100).
- 12 links per GPU, each link with 4 lanes at 50 Gbps.
- Total bidirectional bandwidth: 600 GB/s.
- NVSwitch upgraded to 36 ports.
- Powered DGX A100, connecting 8 GPUs with 4 NVSwitches.
NVLink 4.0 (2022, Hopper H100)
- Implemented on Hopper GPUs (H100).
- 18 links per GPU, 2 lanes each, at 100 Gbps PAM4.
- Total bidirectional bandwidth: 900 GB/s.
- NVSwitch 3.0 with 64 ports.
- DGX H100 architecture: 8 GPUs, 4 NVSwitches, plus integration with 800G optical modules for non-blocking fabric.
NVLink vs PCIe - A Detailed Comparison
From generation to generation, NVLink consistently outpaces PCIe in both per-lane throughput and aggregate bandwidth.
Bandwidth Per Generation
| Generation | NVLink (GB/s bidirectional) | PCIe x16 (GB/s) |
| NVLink 1.0 (2014) | 160 | PCIe 3.0: 15.8 |
| NVLink 2.0 (2017) | 300 | PCIe 4.0: 31.5 |
| NVLink 3.0 (2020) | 600 | PCIe 5.0: 63.0 |
| NVLink 4.0 (2022) | 900 | PCIe 6.0: 121 |
Latency and Scalability
- PCIe relies on switch hierarchies and CPU scheduling → higher latency.
- NVLink provides direct GPU-to-GPU lanes and, with NVSwitch, achieves full connectivity for 8–16 GPUs without CPU involvement.
Use Cases Where NVLink Outperforms PCIe
- AI training: Faster gradient synchronization in large LLMs.
- HPC: Tight GPU coupling in scientific workloads.
- Simulation & rendering: Higher throughput in multi-GPU rendering engines.
The Role of NVSwitch in Scaling NVLink
How NVSwitch Works?
NVSwitch acts as a switching fabric for GPUs, ensuring every GPU in a system can communicate with every other GPU at full NVLink speed.
- NVSwitch 1.0: 18 ports, 50 GB/s each.
- NVSwitch 2.0: 36 ports, 50 GB/s each.
- NVSwitch 3.0: 64 ports, supporting 800G optical interconnects.
DGX Systems Overview
| System | GPU Model | GPUs per System | NVSwitch Version | Bandwidth per GPU |
| DGX-1 | P100 | 8 | None | 160 GB/s |
| DGX-2 | V100 | 16 | NVSwitch 1.0 | 300 GB/s |
| DGX A100 | A100 | 8 | NVSwitch 2.0 | 600 GB/s |
| DGX H100 | H100 | 8 | NVSwitch 3.0 | 900 GB/s |
Future of NVLink and Emerging Technologies
Optical NVLink (Silicon Photonics)
Roadmap boundary: the following preserved optical-interconnect points are historical or forward-looking context. Verify announced products, generation, shipping status, interfaces, reach, topology, software, and support in current NVIDIA documentation before using them in a design.
- Embedding silicon photonics next to GPUs.
- Connecting GPUs via optical fibers for long-distance, high-bandwidth scaling.
- Potential for AI superclusters beyond 256 GPUs.
Integration with InfiniBand and SHARP
Since acquiring Mellanox, NVIDIA is combining NVLink + InfiniBand technologies:
- External NVSwitch chips with SHARP (Scalable Hierarchical Aggregation and Reduction Protocol).
- Enables network-level GPU collectives, reducing bottlenecks in AI and HPC clusters.
Practical Applications of NVLink
- AI Training at Scale: LLMs like GPT, BERT, and diffusion models require thousands of GPUs. NVLink minimizes communication overhead.
- High-Performance Computing: Weather prediction, molecular dynamics, and quantum simulations benefit from lower latency inter-GPU transfers.
- Cloud and Data Centers: Multi-tenant AI workloads rely on NVSwitch-based fabrics for GPU virtualization.
- Financial Services: Faster GPU analytics in real-time trading systems.
Frequently Asked Questions (FAQ)
Q1: What is NVLink?
A: NVLink is an NVIDIA interconnect implemented in specific GPU, CPU, and system generations. Link count, bandwidth, topology, coherency, protocol behavior, and supported peers are architecture- and product-specific.
Q2: How should NVLink be compared with PCIe?
A: Compare the exact PCIe and NVLink generations, lane or link counts, direction convention, topology, peer access, memory behavior, software, message size, concurrency, and workload. A peak aggregate number is not application performance.
Q3: Which NVIDIA GPUs support NVLink?
A: Support varies by exact GPU SKU and form factor; a family name is insufficient. Verify the current GPU and system data sheet, bridge or baseboard topology, firmware, driver, CUDA, NCCL, and peer-access matrix.
Q4: What does NVSwitch do?
A: NVSwitch provides switched NVLink connectivity in supported NVIDIA systems. Port count, bandwidth, nonblocking behavior, GPU domain, management, redundancy, and topology depend on the specific NVSwitch and system generation.
Q5: Will future NVLink systems use optical interconnects?
A: Treat optical NVLink references as roadmap context unless NVIDIA has published a specific supported product and deployment guide. Do not infer availability, reach, topology, price, or compatibility from research or announcements.
Conclusion
Over four generations, NVIDIA NVLink has redefined GPU interconnects, consistently outpacing PCIe in both bandwidth and scalability. NVSwitch has enabled fully connected GPU meshes in DGX systems, while future advancements in optical NVLink and InfiniBand integration may extend scalability to entire AI superclusters.
Selection boundary: NVLink is one component of supported NVIDIA multi-GPU systems. Its value depends on workload communication, memory model, system topology, software stack, scale-up domain, scale-out fabric, operations, power, cooling, lifecycle, and measured results.
Did this article help you or not? Tell us on Facebook and LinkedIn . We’d love to hear from you!
