SuperPi on GPU, were going CUDA

[poke349]

But only benchers care about it. For CUDA to develop it has to have real-world uses that developers can score funding to code for and turn heads to CUDA that way.

It does have its uses and it has a lot of potential. But a lot of people don't like it because it's too "closed" right now (Nvidia-specific).

Programming for CUDA, I would guess, is kind of a nightmare because it's difficult to parallel across so many cores.

Think about it. There's already so much trouble about several cores... but hundreds?

[K404]

Im not a programmer but I agree completely. Allocation and re-assembly across that many "nodes" could be a royal nightmare but I do believe it will need something as big as Photoshop compatability or the ability to crunch Pixar movies before things will REALLY get moving.

Something that will either be used by thousands of people, including a % at professional level, or clients willing to drop millions on something that works

[schmuckley]

Frankenstein Necro-bump for a good idea that never came to fruition.There's gotta be some way to do this;for ATI cards, too.err.."AMD" now

[TimPrice]

Of course, it's called OpenCL.
Go back to the grave.

[_mat_]

Come back to life, sweet thread ...



In the last weeks I've implemented GPUPI, a benchmark that computes pi in parallel via OpenCL. It was something that I wanted to do for years, but never got that far. Somehow I successfully implemented it and it's currently in beta version. Have a look at it, it's pretty fun to crunch pi on your gpu.


GPUPI 1B: AMD Radeon R9 290, NVIDIA GeForce GTX 980 and Intel Core i7-4960X@4 GHz

I am aware that FUGGER and his team wanted to do something like that for some time. I don't want to piss anybody off, on the contrary! I just think that SuperPI on the GPU has something magical and the idea never got out off my head.

I'd also like to dedicate this benchmark to our beloved Turrican. He will always be missed.

Download, technical details & FAQ: GPUPI Beta 1.2

[felix_w]

^ I tried this, this is my result with stock GPU clocks :

[fellix_bg]

So, this is essentially a double precision benchmark? It is unfortunate that many consumer GPUs have artificially limited DP performance.

Anyway, here's an old Fermi:

[_mat_]

So, this is essentially a double precision benchmark? It is unfortunate that many consumer GPUs have artificially limited DP performance.

Some parts use double precision, but most of it is integer.

The calculation itself is split into two parts, that are treated in smaller packages called batches. A batch consists of a number of partial calculations, configured by the Batch Size, and the memory reduction to accumulate all results. The calculation itself uses at least 64 bit integers and doubledouble arithmetic for 500M and smaller. Starting with 1 billion digits each kernel has to make use of additional 128 bit integer routines at a certain point of the calculation. The higher precision is achieved with two unsigned 64 bit integers, but needs more complex algorithms for all basic math functions.

felix_w, nice to see some FirePro scores.

[Buckeye]

WOW I am going to have to give this a try. Thanks !

[felix_w]

_mat_ should we start a new thread, to keep a list with the scores also ?

[_mat_]

Sure, I will start a new one!

Edit: Here it is: http://www.xtremesystems.org/forums/...=1#post5242169

[blindbox]

Hmmm...

Your R9 290 results.


My R9 290x results.


I feel utterly defeated.

[_mat_]

Yes, the 290 seems very strong even tough it has less compute units. I don't own any R9 cards, so I can't really tell you what is going on there. But there must be good reason for this, maybe driver related, maybe a difference in the architecture of GPU.

[FlanK3r]

many thanks man, wow, Radeon is so strong here...

[Lanek]

Single 7970 stock (well 1050mhz )

[Lightman]

My 290X 1050/1325 does 22.102 in Windows 7 Pro x64 With beta Cat. 14.7.

I too was surprised plain R9 290 was faster at a lover clock and tried clock and memory scaling with no surprises, eg. lower clock = slower time.

[_mat_]

This 290X on Haswell-E is currently the fasted card in the bench: http://hwbot.org/submission/2673063_...0x_19sec_690ms

I wonder if the CPU somehow helps achieving the good performance. The whole execution to call the kernels is measured too (well there is'nt much of an alternative), maybe it's pushing the kernels quicker to the gpu.

[STEvil]

What 290x's are you guys using? Some cards with Elpida memory will score poorly due to broken memory clock tables in the bioses and pushing clocks too far will result in the same.

for example my sapphire with stock bios and elpida memory gets about 24 seconds.

[blindbox]

If it's in the BIOS, maybe the memory timings? Lightman mentioned that scaling down the memory clock drops down the results as well. I haven't tested downclocking the memory myself. I'll try it later.

@Mat, hard to say. Those clocks aren't exactly comparable. Let's see, going graphics clockwise:-

core clock 1215/1050 = 1.157
mem clock 1500/1300 = 1.111

assuming 1000 ops,

at 1215 MHz = 50.787 ops/sec
at 1050 MHz = 43.033 ---> that means, 1215 MHz is 18% faster, with a 15% increase in core clock and an 11% increase in mem clock.

We can conclude that the processor does play a small role. I'm on Haswell 3.7 GHz ~ 3.9 GHz (probably running at 3.9Ghz if it's loading the processor single-threadedly)

But doesn't that mean we're still losing out to R9 290? 10% shaders is a lot IMO. Need more tests.

[knopflerbruce]

Nice benchmark Have to give it a shot.

[Lightman]

What 290x's are you guys using? Some cards with Elpida memory will score poorly due to broken memory clock tables in the bioses and pushing clocks too far will result in the same.

for example my sapphire with stock bios and elpida memory gets about 24 seconds.

Yeah, mine is Elpida so this is the differentiator.
When I was mining my R9 290 Hynix card was faster at it with lower clocks than my full R9 290X. I had to run R9 290X @1500Mem to get same hashing speed as Hynix card with stock 1250MHz memory (which didn't OC at all :P).

[blindbox]

Yeah, mine is Elpida so this is the differentiator.
When I was mining my R9 290 Hynix card was faster at it with lower clocks than my full R9 290X. I had to run R9 290X @1500Mem to get same hashing speed as Hynix card with stock 1250MHz memory (which didn't OC at all :P).

I hope it doesn't mean anything in gaming.

[STEvil]

It does and is why some review cards (and most early retail cards with Elpida memory) gave crap numbers even with a "better" fan bios.

[blindbox]

. That really sucks. Quite disappointed with ASUS too, since this is a ROG MATRIX card.

[Mechanical Man]

It does and is why some review cards (and most early retail cards with Elpida memory) gave crap numbers even with a "better" fan bios.

So, is there any way to fix it? Is it bios where those memory tables reside in?

My Asus 290x gave "00h 00m 24.034s PI value output -> 5895585A0"