NVLink 4 Specifications at a Glance
Generation
NVLink 4 (Hopper)
Bandwidth
900 GB/s per GPU
vs PCIe Gen5
7x Faster
Specifications compiled from NVIDIA Hopper architecture whitepapers, Hot Chips 34 presentations, and public technical documentation.
When NVIDIA announced the H100, most of the attention went to the GPU itself: the 80 billion transistors, the Transformer Engine, the FP8 math. But the H100 is not a single chip acting alone. It is a node in a massively parallel system, and the glue that holds that system together is NVLink 4.
If you want to understand why AI models like GPT-4 require entire buildings full of GPUs, you need to understand NVLink. It is the technology that turns a collection of individual processors into a single, coherent supercomputer.
What is NVLink? (And Why PCIe Is Not Enough)
Every computer has a motherboard. On that motherboard, components talk to each other through a standard connection called PCIe (Peripheral Component Interconnect Express). PCIe is a general-purpose bus designed to connect CPUs, GPUs, storage drives, and network cards.
The problem is that AI is not general-purpose. Training a large language model requires GPUs to exchange massive amounts of data—gradients, activations, and model weights—continuously. PCIe, even at its fastest (Gen5), delivers about 128 GB/s. That sounds fast, but it is a bottleneck when your GPU's own memory runs at over 3,000 GB/s.
NVLink is NVIDIA's proprietary solution. It is a direct, point-to-point connection between GPUs that bypasses the CPU and the PCIe bus entirely. It allows one GPU to read and write directly to another GPU's memory.
The Evolution: NVLink 1 to 4
NVLink has been evolving for nearly a decade. Each generation has increased bandwidth and the number of links per GPU. The fourth generation, used in Hopper, represents a major leap.
| Generation | Architecture | Links per GPU | Total Bandwidth |
|---|---|---|---|
| NVLink 1 | Pascal (P100) | 4 | 160 GB/s |
| NVLink 2 | Volta (V100) | 6 | 300 GB/s |
| NVLink 3 | Ampere (A100) | 12 | 600 GB/s |
| NVLink 4 | Hopper (H100) | 18 | 900 GB/s |
NVLink 4 achieves its 900 GB/s bandwidth by increasing the number of links from 12 to 18 and by using faster signaling (50 Gbaud PAM4). Each individual link carries 50 GB/s of bidirectional bandwidth.
How NVSwitch Creates a Full Mesh
Having 18 NVLink ports on a GPU is useful, but it is not enough to connect eight GPUs together in a way where everyone can talk to everyone else at full speed. If you simply wire GPUs directly to each other, you end up with a ring or a mesh where some connections require multiple hops.
To solve this, NVIDIA uses a dedicated switch chip called NVSwitch. The third-generation NVSwitch (used with NVLink 4) has 64 NVLink ports and a total switching capacity of 3.6 Tbps.
In a DGX H100 system, four NVSwitch chips are used to connect eight H100 GPUs. The result is a fully connected, non-blocking topology. Every GPU can communicate with every other GPU at the full 900 GB/s simultaneously.
NVLink vs. PCIe vs. InfiniBand
These three technologies are often confused, but they operate at different layers of the system.
- NVLink: The connection inside a server, between GPUs. It is the fastest and lowest-latency link.
- PCIe: The connection between the CPU and the GPU (and other components). It is a general-purpose bus, not optimized for GPU-to-GPU traffic.
- InfiniBand / Ethernet: The connection between servers. It is slower than NVLink but can span much greater distances, connecting thousands of nodes in a data center.
| Technology | Bandwidth | Use Case |
|---|---|---|
| NVLink 4 | 900 GB/s | GPU-to-GPU (inside a server) |
| PCIe Gen5 x16 | 128 GB/s | CPU-to-GPU, general I/O |
| InfiniBand (400G) | 50 GB/s per port | Server-to-server (across a data center) |
The Secret Weapon: In-Network Computing (SHARP)
NVLink 4 and the third-generation NVSwitch introduced a feature called SHARP (Scalable Hierarchical Aggregation and Reduction Protocol). This allows the switch itself to perform parts of the all-reduce calculation as data passes through it.
Instead of sending all the data to every GPU and having each GPU do the math, the switch can aggregate the data and send only the result. This reduces network traffic and significantly speeds up training. NVIDIA claims a 2x increase in all-reduce throughput compared to the previous generation.
Why NVLink 4 Matters for the H100
The H100 is designed for massive AI clusters. Without NVLink 4, the H100 would be limited by PCIe, and the performance of large-scale training would be severely bottlenecked. NVLink 4 is what makes it possible to treat eight H100 GPUs as a single, unified compute engine.
It is also a key enabler for Unified Memory in Grace Hopper superchips, where the CPU and GPU share a coherent memory space. The same philosophy of fast, direct memory access that powers NVLink 4 is now being pushed down to consumer devices like the RTX Spark.
The Bottom Line
NVLink 4 is not a flashy product. It does not have a marketing page filled with game benchmarks. But it is one of the most important pieces of technology in the modern AI stack. It is the difference between a pile of GPUs and a supercomputer. As AI models grow larger, the importance of the interconnect will only increase, which is why NVIDIA is already pushing ahead with NVLink 5 and NVLink 6 for its Blackwell and Rubin platforms.
Frequently Asked Questions
Can I use NVLink on a consumer GPU?
Some older consumer GPUs (like the RTX 3090) supported NVLink for linking two cards together. However, NVIDIA has removed NVLink support from recent consumer cards (RTX 40 and 50 series) to differentiate them from professional and data-center products. NVLink is now primarily a data-center feature.
What is the difference between NVLink and NVSwitch?
NVLink is the physical connection (the cable and protocol) between two GPUs. NVSwitch is a chip that acts as a central hub, allowing multiple GPUs to connect to each other through it. NVLink is the road; NVSwitch is the intersection.
Does NVLink replace PCIe?
No. They serve different purposes. PCIe connects the CPU to the GPU. NVLink connects GPUs to each other. A server needs both.
Sources & Fact-Check
This article was drafted with AI assistance and reviewed by a human editor. All specifications, pricing, and performance figures were cross-referenced against the following primary sources: