Nvidia and AMD have spent many years trying to perfect their multi-GPU configurations, namely SLI and crossfire. PC Enthusiasts have always supported both the teams and stuffed their cases full of GPUs in hopes that they will not face diminishing returns.
However, due to a lot of work required by game developers and GPU manufacturers with no chance of mass adoption in sight, SLI and crossfire have been pushed too low priority for many years, and enthusiasts started to stick to one single but powerful graphics card in their machines.
Nvidia, however, has started bringing SLI from the dead by introducing NVlink, a new solution for the people who want to rock multi GPUs and try to experience next-generation levels of performance right now. So, let me explain What NVLink is and what you need to know about it.
What Is NVLink, And How Does It Work?
If you are a PC gamer, then you might have heard the term SLI which means using two or more NVIDIA graphics cards on a PC simultaneously. The two graphics cards were connected using an SLI bridge, which let both cards communicate with each other.
In SLI, one card worked as the master card while the other (s) worked as slaves. The slave cards rendered the frame while the master did the rendering and compilation of the frames to be sent to the monitor.
The memory buffer of the master card was the only one used, which made the additional memory effectively useless during the workload.
NVLink, on the other hand, tries to solve the problems of SLI. Nvidia’s Multi-GPU branding remains SLI; however, the SLI bridge that connects the GPUs has been replaced by NVLink.
In NVlink, the two cards, when connected, no longer have a hierarchical structure as they form a mesh network. This means that there are no master or slave cards anymore.
Both the cards work simultaneously to perform the workload, and the resources of both the cards get pooled together. So, the memory bus of both the cards is also fully utilized.
However, NVlink is mainly targeted towards enterprises, and consumers NVLink still faces the same gripes which SLI has faced, like no interest from developers and extremely low adoption. Interest was slowly going up in SLI when NVlink was introduced; however, it died down just as soon.
SLI Vs NVLink In 2023: What’s The Best Multi GPU Configuration For You?
Since the days “3dfx” was a famous graphics card manufacturer, SLI has been with us. Nvidia bought the technology from then and started implementing it in their GPUs.
For some time, SLI was popular with the enthusiasts; however, due to low adoption and not enough support from the developers, it started to die down.
In SLI, graphics cards were placed in a Hierarchial structure where one card acted as the master while the other served as slaves.
This meant that one card did the bulk of the work and used its resources while the other only rendered alternative frames with one-way communication of only 2 GB/s.
This caused diminishing returns, and adding more cards did not mean more performance which effectively killed SLI in the consumer space.
NVlink emerged as the successor to SLI and tried to solve the problems which plagued it. The first thing NVlink did was stop using the hierarchical way of linking cards together and instead joined them together using something called a mesh network.
This allowed the cards to pool their resources and talk to the CPU. Two-way communication was enabled. NVlink was also very fast as it could transfer data at 150 GB/s.
Now, if we were to take stock of it, I believe NVlink is the way to go; however, there are some caveats. Consumer NVLink is limited to only two cards, while enterprises can use at least 16 cards at once.
There is still the problem of low adoption and lack of support from developers and interest from NVidia themselves. So, I believe that it is better to stay away from Multi-GPU configurations and stick to one single but powerful GPU.
Diminishing returns happen in Multi-GPUs as even two RTX 3090s running through NVlink takes a lot of power to run, making your PC effectively a space heater, but only gives you a slight bump in performance making it a waste of money for the most part.
What Is NVSwitch?
NVswitch is a technology that PC gamers, I believe, will never use; however, if you are a data scientist or someone working on AI technology, then you might have heard about it.
NVSwitch is a connector that allows enterprise-level supercomputers to connect 16 GPUs already connected using NVLink and effectively scale their workloads together to do deep learning, data mining, and other tasks.
Supercomputers like the DGX always need more performance to complete their parallel processing, and NVlink can connect a set of 8 GPUs at one time. The NVswitch combines two sets of these 8 GPUs, making them 16, which gives out an exceptional level of performance.
How NVLink And NVSwitch Work Together?
Suppose you are working on a project that needs a lot of performance and your Quadros at home are not enough to do the task.
You use a supercomputer and are using GPUs like the H100 from Nvidia, and you see that there are a lot of them. These GPUs work together using NVlink and NVswitch.
There are eight H100 GPUs connected and working together using NVlink. But you find yourself low on processing power and require another set of eight H100 so that you can do your work. But how do you connect these two sets of NVlinked GPUs? Well, that’s where NVswitch comes in.
In enterprises, NVLink and NVswitch are used simultaneously and go hand in hand. The NVswitch allows you to connect these two sets of NVlinked GPUs and work together.
They form the same mesh you wanted and effectively give you the power of 16 GPUs simultaneously.
What Graphics Cards Support NVLink?
The following Nvidia graphics cards support NVlink.
- Geforce RTX 3090
- Titan RTX
- Geforce RTX 2080Ti
- Geforce RTX 2080
- RTX A6000
- RTX A5000
- RTX A5500
- RTX A4500
- Quadro GP100
- Quadro GV100
- Quadro RTX 5000
- Quadro RTX 6000
- Quadro RTX 8000
Please keep in mind that there are different generations of NVLink available, and not every NVlink connector will be compatible with every graphics card.
Is NVLINK Worth It?
The answer to the question about NVlink being worth it or not is that it depends. NVlink is undoubtedly a great attempt at bringing SLI and multiple GPU configurations back into vogue.
However, if you are a PC gamer, it would be better to stick with one graphics card instead of NVLink. This is because of game developers’ general lack of interest in optimizing their games for dual GPUs.
So, the diminishing returns make NVlink a waste of money, and SLI is effectively dead in the consumer market.
However, if you are someone that works with supercomputers and does AI and deep learning, the NVLink is one of the best things ever. It will allow you to pool all the resources of your graphics cards and give you all the processing power you need.
Nvidia has been one of the best GPU manufacturers for over a decade now, and gamers like myself will always support them by buying their graphics cards.
However, in the consumer space, they have missed the mark regarding NVlink and Multiple GPU solutions.
NVLink link is excellent at the enterprise level, and I do suppose data scientists no longer have their work cut out for them as they now have access to so much raw horsepower to do what they need to do.
There is no doubt always room for improvement, but the NVlink and the NVswitch are a godsend for the enterprise.
Frequently Asked Questions
Yes! NVlink creates a mesh network of GPUs to pool their resources and effectively add the VRAM to their arsenal.
You can use NVlink to connect only two GPUs in the consumer space. However, when it comes to enterprises, you can combine 16 enterprise-grade GPUs using NVlink and NVswitch.
#3 – Does NVLink increase VRAM?
NVlink does not increase your VRAM; however, it does the pool and adds together the video memory inside your GPUs. This was impossible in SLI as only the Vram of one GPU was used.
On the enterprise side of things, NVLink is constantly getting improved. However, NVlink faces a lack of interest from Nvidia in the consumer space. NVlink can still be used on the RTX 3090.
The Nvidia Geforce RTX 3090, Titan RTX, Geforce RTX 2080 Ti, Geforce RTX 2080, RTX A6000, RTX A5000, RTX A5500, RTX A4500, Quadro GP100, Quadro GV100, Quadro RTX 5000, Quadro RTX 6000, Quadro RTX 8000 all support NVlink.
#6 – Does VRam stack in SLI?
VRAM does not stack in SLI due to the master-slave configuration. The VRAM of only the master card is used in SLI.
An NVlink will cost you around $80 to $100.
Yes, Nvidia still supports Nvlink; however, it is currently limited only to the RTX 3090 and other enterprise-grade graphics cards.