GPU programming and benchmarks

[martward]

Hello all,,

 

I am going to do my master thesis next year, and I am probably going to do a project about Information retrieval. Since this involves a lot of data, "regular" programming is not going to suffice for a big project. Therefor I want to learn how to program using the CUDA cores on my GTX 950. Is there anyone who has learnt this and who can tell me what a good way/site/book to learn this is?

 

Also, would it be an idea to include some benchmark for GPU programming on big data to GPU reviews, especially since NVIDIA's new architecture is even advertised with the improvements for GPU programming for big data and artificial intelligence applications? I am thinking about trading my gtx 950 in for a Pascal GPU (once the budget models are out) but I have no idea where I could find information about the performance of the cards in these kind of scenario's. Maybe we can force Luke to cover it when he reviews the Pascal GPU's.

 

Thanks

[trag1c]

CUDA is fairly straight forward. Allocate memory on the device, upload to that memory, invoke the function with x number of threads with y number of thread groups, check if execution is done, then transfer results back to the host.

 

I learned CUDA through the documentation thats available on Nvidia's developer website so its probably a good place to start.

[maremp]

On 3/12/2016 at 0:30 AM, martward said:

Maybe we can force Luke to cover it when he reviews the Pascal GPU's.

Even if you were to "force" (imo bad choice of words, 'ask' sounds nicer) him, they would probably do the flagship gpu, not the lower-end one, especially for cuda performance, which is relevant to prosumer market. 

[martward]

4 hours ago, maremp said:

Even if you were to "force" (imo bad choice of words, 'ask' sounds nicer) him, they would probably do the flagship gpu, not the lower-end one, especially for cuda performance, which is relevant to prosumer market. 

Ofcourse I didn't mean I was going to fly to Canada to chase him with a pointy stick until he made a video about it :p. I meant that if there would be enough interest in the community we could do like a poll or something to encourage them to do it. I was thinking of just an extra benchmark instead of doing a completely seperate video, that way it would be easy to test it for all video cards. I don't think there is a very broad audience for this though.

[maremp]

I understand what you've meant and most users would probably love to see something more low-end. From marketing point of view, they are most likely not going to do a 30 min long video with wasting tens of hours on benchmarking just for this. And nvidia is not going to support this either, it's in their interest to advertise how awesome their flagship gpu is, they aren't going to invest much on promoting something like gtx 950.

But it's sad how the channel is becoming more of a business venture and balls to the wall build guides that maybe 0.001% viewers can afford. I too like old videos where they reviewed stuff that I would actually consider buying and I'm not particularly interested in cutting a $1000 gpu in half. 

[JustJollyJonah]

I agree with maremp but with the current amount of staff LMG has employed it's just not profitable for them to review lower end Cards.

Mostly this is due to the retailers, MSI and others want to show how great their new high-end card is and don't want to waste any money on marketing lower end cards that have the performance from previous generation with minuscule improvements.

And as Linus has previously made a few videos on, it's just not a interesting card since there has been no innovation in the lower end cards for some time.

 

As for CUDA programming, it's really not that hard too learn, just google some guides, or go on stack exchange and look for some code examples.

This PDF has some great examples: CUDA Programming guide