Why would I want multiple GPUs?

[Crazy Mad Dog Blade]

I'm under the impression that GPUs only process graphics. I also heard that GPUs can be used to computate non-graphics stuff. Can anyone elaborate? 

[Fendrick]

I dunno, why would you want multiple GPU's? Can you elaborate?

[Crazy Mad Dog Blade]

1 minute ago, Fendrick said:

I dunno, why would you want multiple GPU's? Can you elaborate?

Cases often have multiple slots for multiples GPUs and when I talked to a company that built super computers, they asked me how many gpus I wanted in one build. 

[Budget DIY]

2 minutes ago, Crazy Mad Dog Blade said:

Cases often have multiple slots for multiples GPUs and when I talked to a company that built super computers, they asked me how many gpus I wanted in one build. 

Multiple GPU's is called "SLI" (on nvidea cards), and it's much less common.

 

Some people can use multiple GPU's, to do *magic*. (AI stuffs, computer learning and whatnot) More common to have multiple GPU's, is crypto mining.

 

If you plan on just playing games and do other random consumer grade stuffs, you will be fine with a decent GPU, which fits your budget.

[BondiBlue]

That depends on what you're wanting to do with your computer. I have a few computers that each have multiple GPUs installed, but they're not used together for gaming. Instead I take advantage of the additional GPUs by using them for Folding@home, which uses the GPUs for computational power. Some people have multiple cards installed for mining purposes, and others use them simply to add additional display outputs. 

[Dr. Will0hlep]

In short, GPUs are good at Linear Algebra (maths but with matrices) and while this makes them good at graphic, it also makes them useful for modern machine learning work which is almost all linear algebra. GPUs also tend to have the tools for video encoding which makes them useful for streaming and video editing.

 

Generally the main reason to have lots of them is if you need to do alot of machine learning work or crypto mining. I imagine there are other reasons but these are the common ones.

[Eigenvektor]

17 minutes ago, Crazy Mad Dog Blade said:

I'm under the impression that GPUs only process graphics. I also heard that GPUs can be used to computate non-graphics stuff. Can anyone elaborate? 

GPUs are good at specific types of math. While they were purpose built for graphics in the past, turns out being able to perform thousands of matrix multiplications in parallel is good for other use cases (e.g. physics, AI, …).

 

Multiple GPUs for gaming is pretty much a thing of the past. It never worked all that well and modern games have abandoned SLI/Crossfire. However non-gaming workloads like AI can still benefit from having multiple GPUs.

 

20 minutes ago, Crazy Mad Dog Blade said:

when I talked to a company that built super computers, they asked me how many gpus I wanted in one build. 

By super computer are you talking about actual supercomputers used by universities and the military, or some place that markets their gaming machines as "super computers"?

 

An actual supercomputer is generally used to run scientific or military simulations and often contains thousands of CPU/GPU cores. The computer may be running hundreds of different projects at once, each using one or more CPUs and/or GPUs in parallel to run tons of calculations in parallel.

[Crazy Mad Dog Blade]

6 hours ago, Eigenvektor said:

GPUs are good at specific types of math. While they were purpose built for graphics in the past, turns out being able to perform thousands of matrix multiplications in parallel is good for other use cases (e.g. physics, AI, …).

 

Multiple GPUs for gaming is pretty much a thing of the past. It never worked all that well and modern games have abandoned SLI/Crossfire. However non-gaming workloads like AI can still benefit from having multiple GPUs.

 

By super computer are you talking about actual supercomputers used by universities and the military, or some place that markets their gaming machines as "super computers"?

 

An actual supercomputer is generally used to run scientific or military simulations and often contains thousands of CPU/GPU cores. The computer may be running hundreds of different projects at once, each using one or more CPUs and/or GPUs in parallel to run tons of calculations in parallel. 

By supercomputer, I meant xenowulf. That was the company I was talking to and asked how many gpus I wanted in the build. Thank you for your help!

[Eigenvektor]

4 hours ago, Crazy Mad Dog Blade said:

By supercomputer, I meant xenowulf. That was the company I was talking to and asked how many gpus I wanted in the build. Thank you for your help!

In a sense their website already answers your question:

Quote

Custom-built, liquid cooled, multi-GPU workstations, powerful enough to take on the most intensive rendering and deep learning workloads.

Rendering in this case refers to things like rendering movies (e.g. CGI)

 

This is what I have in mind when I hear supercomputer:

 

[Coaxialgamer]

You can quite easily slot multiple GPUs into a single rig. Whether that's worth it to you will depend on your usecase.

 

Many compute application will see those seperate processors and have no issue exploiting them: 3D rendering, mining, and other compute-heavy application can benefit from multiple cards. 

 

That's still not going to work well for software that can't explicitly take advantage of multiple GPUs though, such as games (with the exception of DX12 LDA explicit mode). For that you historically had something called SLI or Crossfire on the Nvidia and AMD/ATI sides respectively. Typically you had a small cable linking cards together, and that would allow potentially up to 4 cards to work on something like a game, assuming it was supported by the devs.

 

Nvidia also has something called Nvlink which is similar and allows cards to pool ressources and share memory (basically more advanced SLI), but that's mostly reserved for server-side hardware and doesn't exist on current consumer stuff.

 

For games, both Nvidia and AMD have now stopped supporting multi-GPU entirely, seeing as it often didn't work very well and the value proposition was always relatively poor.

For compute you can still do either of the other things I mentioned. I should note that you do absolutely NOT need SLI to work using two or more GPUs.

[Crazy Mad Dog Blade]

19 hours ago, Eigenvektor said:

In a sense their website already answers your question:

Rendering in this case refers to things like rendering movies (e.g. CGI)

 

This is what I have in mind when I hear supercomputer:

 

Yep not that lol. Do you know of any alternative services to xenowulf that would build a rig with a AMD 5995wx?

[igormp]

1 minute ago, Crazy Mad Dog Blade said:

Yep not that lol. Do you know of any alternative services to xenowulf that would build a rig with a AMD 5995wx?

Many workstation integrators provide those, such as:

https://www.pugetsystems.com/workstations/threadripper-pro/wrx80-e/

https://shop.lambdalabs.com/gpu-workstations/vector/customize

https://www.digitalstorm.com/configurator.asp?id=4517852

 

Lenovo also have their ThinkStation lineup, and so does Dell with their precision workstations.