Can I use a GPU as second CPU (at least for graphics tasks) ?

Hello!

Probably a stupid question, probably has been asked before, but I was wondering. If you have a discrete graphics card, can you use its power for common tasks? I know that if you have nVidia for example, and the application supports Cuda, you get faster speeds than if you disable Cuda and all the work is done by the CPU. My question is if I can force my nVidia GPU to help for applications that do not support Cuda, at least for image or video processing (such as VirtualDub). Is there any way to make the GPU appear as extra CPU cores/threads so it helps with processing? I could force an application to use the GPU cores by manually changing the task affinity. Is there any other similar trick? Please don't judge, I already know this might be a stupid question, but I ask it just in case it can be done and I'm missing something. My humble (by today standards) Core-i7 3770 could benefit from my nVidia GTX 1050 Ti help. Why not?

Thank you in advance.

spapakons said:

Hello!

Probably a stupid question, probably has been asked before, but I was wondering. If you have a discrete graphics card, can you use its power for common tasks? I know that if you have nVidia for example, and the application supports Cuda, you get faster speeds than if you disable Cuda and all the work is done by the CPU. My question is if I can force my nVidia GPU to help for applications that do not support Cuda, at least for image or video processing (such as VirtualDub). Is there any way to make the GPU appear as extra CPU cores/threads so it helps with processing? I could force an application to use the GPU cores by manually changing the task affinity. Is there any other similar trick? Please don't judge, I already know this might be a stupid question, but I ask it just in case it can be done and I'm missing something. My humble (by today standards) Core-i7 3770 could benefit from my nVidia GTX 1050 Ti help. Why not?

Thank you in advance.

Click to expand...

Hi there
I don't think this is possible as many internal native CPU machine code instructions wouldn't exist on the GPU.

However with emulation and Virtualisation -- who knows if it's theoretically possible, The question is for example could you use a hose pipe as electrical wire. Different hardware for different tasks. So i'd imagine no -- also these things are designed to operate as co-processors not as primary processors.

Cheers
jimbo

A CPU is designed for general purpose computing, a GPU on the other hand is designed for specific tasks repeatedly done in parallel. Most software are not designed to do tasks in such a way. Thus it's impractical to even attempt to route such tasks to the GPU for software that can't take advantage of concurrent calculations. Also a single GPU core/unit/thread is a lot slower than a CPU core. The speed on a GPU comes from the huge number of parallelism.

Well, some video editing applications, including VirtualDub, support DirectX hardware acceleration. Question: If I use this, does the GPU help, or it is done from the CPU? It there a DirectX to Cuda wrapper to pass these tasks to the GPU or compute them faster?

spapakons said:

Well, some video editing applications, including VirtualDub, support DirectX hardware acceleration. Question: If I use this, does the GPU help, or it is done from the CPU? It there a DirectX to Cuda wrapper to pass these tasks to the GPU or compute them faster?

Click to expand...

Yes, you can use GPU assisted video encoding. That's a good use case.

There are a few applications that are written such that they may use some routines within the GPU.

"Why not?"
You cannot just force any application to do that. The cores in a GPU are much developed for a specialist task. The CPU cores/threads/performance cores/efficiency cores for a much wider range of general tasks.

The client for Stanford's "Folding @ Home" allows you to "use" the CPU or GPU or both.
But I believe that's built into the client itself.

The point is that you can only use the GPU for computation if the application is specifically written for such operations.

GPUs are good at crunching numbers in the masses, but that is pretty much all it can do.

If an app is used for such tasks, then there's probably already built in code that can utilize the GPU, otherwise it's pointless as the app would just run slower.

SlicEnDicE said:

If an app is used for such tasks, then there's probably already built in code that can utilize the GPU

Click to expand...

Yep.



But... if an app can be coded to utilize the GPU as well... then it should be possible to create an app that can do that for a program that is run through that created app. Sort of like an emulator or container?

I have a feeling that DirectX acceleration doesn't necessarily mean the GPU does all the work. The CPU is still involved if not doing everything. A DirectX to Cuda wrapper would convert DirectDraw or Direct3D commands to Cuda code forcing an nVidia card to do all the work, hopefully faster.

spapakons said:

I have a feeling that DirectX acceleration doesn't necessarily mean the GPU does all the work. The CPU is still involved if not doing everything. A DirectX to Cuda wrapper would convert DirectDraw or Direct3D commands to Cuda code forcing an nVidia card to do all the work, hopefully faster.

Click to expand...

It does not work that way...DirectX already does (Hardware acceleration) what is required straight on the GPU for the parts that makes sense to do there, no need to convert anything.

In the digital photography area, software that uses AI makes heavy use of the GPU for processing.

Examples: the Topaz AI products, DXO Photolab, and some of the AI filters in Adobe Photoshop.

I've never used them, but some video processing software can use multiple GPUs. DaVinci resolve is one. Would you like to own a PC with two RTX 4090 cards, at 1.6k USD or more each?

GPUs were used for cryptocurrency mining. (Specialized computations.) That led to a shortage of high-end graphics cards. A couple of years ago, most of the cards were only available from scalpers (resellers).

nVidia is now selling accelerator cards (with no graphics ports) for AI. I don't know whether those are their main revenue source now or not. Such cards have existed for years, but the market has picked up.

My CPU, Intel Core-i7 3770 has onboard graphics (Intel HD 4000), but my motherboard (Asus P8H61) is a full ATX without the display outputs. Is it possible to use the Intel graphics in cooperation with the discrete nVidia graphics, or the Intel is only activated on motherboards with display outputs? Of course in UEFI firmware (BIOS) there is no way to activate the onboard graphics, can be done by software?

Yes that's possible in theory. I can't recall exactly how on your particular setup though. It can be a bit finicky to find the correct drivers though that plays well with the other hardware for best results.

I doubt my Intel HD 4000 would make any difference though since my nVidia GTX 1050 Ti would have very little boost from its cooperation. Thank you guys.

Both VirtualDub and VirtualDubMod were abandoned many years ago. It's been over a quarter of a century since I first started using VirtualDub (on a Pentium III @450 MHz), and the last time when I used VirtualDubMod was more than 15 years ago (on a Pentium IV @3.4 GHz). Back then Adobe Photoshop didn't even support CUDA acceleration yet, as Photoshop 5 hadn't come out yet.

The Core i7-3770 supports Intel Quick Sync which is a separate core inside the CPU. Quick Sync is an Intel-specific CPU hardware feature that is available only on specific CPU models from Intel. This feature is dedicated so its only purpose is hardware accelerated video encoding in apps that can support Quick Sync, e.g., Adobe Premiere. It offers a dramatic speed improvement to video encoding in Adobe Premiere, but sacrifices some quality in the video result. (As an aside, Intel Arc Graphics adds hardware accelerated encoding of AV1 video.)

When you have a discrete GPU like the GTX 1050 Ti the main important benefit of also having the HD Graphics 4000 iGPU (integrated GPU) inside the CPU (besides the Quick Sync feature) most commonly is the efficiency benefit. It saves power and generates less heat during those times when the much heavier processing power of the discrete GPU is not required for anything in which case the discrete GPU remains inactive, as the iGPU takes over the graphics related workload from it until this workload becomes too high for the iGPU to still be able to keep up. Intel iGPU processing power on older CPU models from Intel is notably very slow when compared to any decent discrete GPU from around the same era. Although this still holds true for newer CPU models, Intel Iris Xe Graphics (in Intel 11th Gen CPUs) has been a giant leap forward nevertheless, as there is no comparison between Intel HD Graphics 4600 (in Intel 10th Gen CPUs) and it.

SlicEnDicE said:

The point is that you can only use the GPU for computation if the application is specifically written for such operations.

GPUs are good at crunching numbers in the masses, but that is pretty much all it can do.

If an app is used for such tasks, then there's probably already built in code that can utilize the GPU, otherwise it's pointless as the app would just run slower.

Click to expand...

Some very specialized Floating point numeric processors work too -- actually the idea of those is quite old -- back even in 30386 / 40386 CPU days you could get an F.P co-processor -- with the speed (???) of processors in those days if you did any engineering calculations you'd have need of those.


Any sort of decryption software which was encrypted using several million digit prime numbers would probably have taken longer than the entire probable future lifetime of the universe to get even 1% of the way through its task. (Even now that type of thing requires massive compute power --probably why this year or shortly afterwards Quantum computing finally moves from the lab into some sort of field trial.

cheers
jimbo

Yes, I know the original VirtualDub is too old, but it has some very handy filters, so I still use it alongside the newer VirtualDub 2 version. With the appropriate codecs (x264vfw and AC3ACM) it can create AVCHD compatible AVI video (AVC MPEG4 with Dolby Digital Stereo sound) which can be imported to MultiAVCHD to create an AVCHD-DVD disc without re-encoding and affecting quality. So I get the benefit of VirtualDub's filters in a modern format. I can also create x265 video with the x265vfw codec. VirtualDub 2 can even encode in 10-bit color.

GPUs have dozens and even 100+ cores specifically made to crunch numbers and work in parallel as 3D animation is incredibly complex so their potential for math solving problems is greater than CPUs which also have dedicated FPU. Ai is going to make use of those.

AI has been crunched on GPUs for ages already, because as stated before, GPUs are good at crunching floatingpoint and integer numbers in parallel and that's basically all AI as we know it today is...only diff in the furure is that GPUs will get dedicated AI cores. Even next gen. Mobile phones will have some pretty advanced and impressive AI hardware built in. Mobiles already use GPU accelerated or dedicated AI chips for image processing in cameras.