CUDA and openCL Benchmarks

Titan
Titan
Joined: 29 Aug 13
Posts: 19
Credit: 25868802
RAC: 0

Sid I have found that Perseus

Sid I have found that Perseus BRP task require little memory per task, running 3 at a time only consumes 680-700 mb. This changes slightly depending on which driver is being used. I currently am using 2 R9 280x, the Sapphire Toxic model with 3gb vram. I have found 3 tasks per card to be ideal for my setup(using the 13.12 drivers) however this is with a z77 chipset which is not ideal for multiple cards due to the limited PCI e lanes. I have found pci e bandwidth seemes to be the limiting factors for this task.

David Rapalyea
David Rapalyea
Joined: 3 Jan 13
Posts: 79
Credit: 63886821
RAC: 0

Question I notice that my

Question

I notice that my GTX 660 has about twice as many CUDA cores as my GTX 650 cards and the GTX 670 only has about another 50%. Am I correct in judging our tasks are CUDA core intensive and GTX 670 would not be as good a choice as the other two based on cost?

I am thinking about picking up a couple of EBAY GTX 660 cards for my two-card machine and replace the GTX 650 cards now in the machine and suspect I should not mess with GTX 670 cards.

Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go

tbret
tbret
Joined: 12 Mar 05
Posts: 2115
Credit: 4862880478
RAC: 139333

RE: Question I notice that

Quote:

Question

I notice that my GTX 660 has about twice as many CUDA cores as my GTX 650 cards and the GTX 670 only has about another 50%. Am I correct in judging our tasks are CUDA core intensive and GTX 670 would not be as good a choice as the other two based on cost?

You didn't say what the cost is, so it's really hard to say.

I find that my 670s are about 15% faster at BRP4 tasks than my 660Ti cards. It is important to note that both of those machines are equal, but not optimal. The cards are running at x8 on PCIe2, instead of x16 on PCIe3. It's really not possible for me to extrapolate what their relative performance would be on your system.

My 560s (no Ti) at x16 are actually faster than my 660Tis at x8.

Go ahead and hate me for saying it; if all you crunch is Einstein, you really owe it to yourself to try a R9 AMD card.

My R9 270X on PCIe3, x 16, w/1800MHz RAM, is 30% faster than my 670s at x8 w/1334MHz RAM. I have already ordered some 1866MHz RAM for that system and will see how much of that is RAM speed. I should be installing it tomorrow night.

The R9 was also a lot cheaper than my latest NVIDIA purchase, which was a 770, which runs only slightly faster than the 670s. (the AMD cost just more than half as much)

But that's only true with the program here at Einstein. Elsewhere the 670 and 770 are faster than the R9 270X, and the PCIe and RAM speed is less critical.

If it were *me*, I wouldn't buy any more new NVIDIA cards until the new models arrive, and if I were going to buy used cards I'd go all the way back to the 560 and save whatever money I could.

Why?

I have two old 470s and a 560Ti-448, and a 560Ti OC model in one case. It produces about 40% more than the quad 660Ti machine and about the same as my 670/770 based machine.

BUT - the i5 machine with the one R9 in it? It produces about 40% of what the fastest of my quad-card NVIDIA machines produce.

But that's my machines, partially hampered by RAM speed and partially hampered by PCIe 2.0 slots running at x8.

I don't think that changing a 670 or 770 into a PCIe 3 slot is going to make-up the performance difference, but I'm quite sure it will not make-up the price and performance differences combined.

MAGIC Quantum Mechanic
MAGIC Quantum M...
Joined: 18 Jan 05
Posts: 1886
Credit: 1406604598
RAC: 1171231

RE: Go ahead and hate me

Quote:

Go ahead and hate me for saying it; if all you crunch is Einstein, you really owe it to yourself to try a R9 AMD card.

nVidia will hate you tbret ...........if you get me to switch to this one.

Is this the one?

Now my only question is will this work better with my older 3-core phenom?
I switched the XP Pro to Win7 so I can use the 8GB ram now.

http://einsteinathome.org/host/11652712

I had that nVidia 650Ti OC in that one before but put that in another host while this one was down.

The price looks good especially with the 1280 Stream Processors

It would be my first one that isn't a GeForce

tbret
tbret
Joined: 12 Mar 05
Posts: 2115
Credit: 4862880478
RAC: 139333

RE: RE: Go ahead and

Quote:
Quote:

Go ahead and hate me for saying it; if all you crunch is Einstein, you really owe it to yourself to try a R9 AMD card.

nVidia will hate you tbret ...........if you get me to switch to this one.

Is this the one?

Now my only question is will this work better with my older 3-core phenom?
I switched the XP Pro to Win7 so I can use the 8GB ram now.

http://einsteinathome.org/host/11652712

I had that nVidia 650Ti OC in that one before but put that in another host while this one was down.

The price looks good especially with the 1280 Stream Processors

It would be my first one that isn't a GeForce

Well, the two I have are these:

[url]http://www.bestbuy.com/site/xfx-radeon-r9-270x-2gb-ddr5-pci-express-3-0-graphics-card/2091002.p?id=1219068840893&skuId=2091002&st=categoryid$abcat0507002&cp=2&lp=14[/url]

... what can I say? They were there.

The Phenom 3 core? Yes, it'll work, but it is going to cost you something like 20% of the performance.

Here's what I am seeing in my own machines. A Phenom II 1090T with 1334MHz RAM, with a R9 270X, PCIe v2.0, x16 is running 20% or more SLOWER than the i5 with a PCIe 3.0 x16 and 1800MHz RAM.

But because of some other things I see in other machines, I have a feeling that most of the difference is the increase in the RAM bandwidth. But to get 1866 RAM to work in that computer, I have to replace the 1090T with a FX processor.

I bought the 1866MHz RAM to try in my machine with the 670s in it, which does have an FX processor in it, to see how much difference the faster RAM makes.

But you really do have to forgive me for not being curious enough to start disassembling machines and reassembling them.

I'm still playing.

I don't hide my computers, so you're more than welcome to look around in their results. I'll be happy to tell you whatever I can about any of them.

Let's see what happens when I replace the 1334 RAM with the 1866 RAM. Like I said, sometime tomorrow night. If there is no speed difference in the GTX 670 system, the it would be a huge waste to upgrade all the RAM and all of the processors, etc.

Edit: By the way --- If I were going to buy one card and that's all, I'd have to think long and hard about some of the R9 280 cards (notice; no X) to get "as close" to the real thing as I would afford.

My machines are so varied... it's really, really hard to generalize anything by looking at them. One step at a time. RAM is next.

EDIT the EDIT: The machine that I just don't really understand is the Phenom II 965. That's two OC 560s in X16 slots and one OC 560Ti in a x4 slot, all of it running with 1334 RAM. I don't see much difference (if any) between the 560s and the 560Ti. Soooo, is it the x4 slot? Is it that the cards are all choked by the "lanes" or is it the RAM or... what?

There are too many possibilities for me to mess with it all. I'm looking for something I can do that isn't expensive that gets me the best output I can get out of what I've got without replacing it all.

THEN, I'm going to maximize what I can, then start shutting some machines down. The electricity bills are obscene.

MAGIC Quantum Mechanic
MAGIC Quantum M...
Joined: 18 Jan 05
Posts: 1886
Credit: 1406604598
RAC: 1171231

RE: THEN, I'm going to

Quote:
THEN, I'm going to maximize what I can, then start shutting some machines down. The electricity bills are obscene.

No doubt about that!

I never even checked to see how much my 7 hosts cost to run 24/7 until I read other members here talking about that.....soooo.......when I did check myself a couple months ago I was a bit surprised and not in a good way.

Especially if I thought about how I have been doing this for 15 years.

Good thing we didn't start with Seti Classic and GPU tasks.

(I am going to try to keep myself from doing the math of just how many KW of power I have used since I started doing these GPU tasks)

But I still would like to put a card in my host #7 just because......and of course the most bang for the buck

So a fuel efficient card for the best price.

(I also run the vLHC,LHC,amd Atlas tasks)

Ok its after midnight.......so thanks........and goodnight tbret

mikey
mikey
Joined: 22 Jan 05
Posts: 12680
Credit: 1839083724
RAC: 3916

RE: Go ahead and hate me

Quote:

Go ahead and hate me for saying it; if all you crunch is Einstein, you really owe it to yourself to try a R9 AMD card.

Maybe and maybe not...
tbret is using his AMD Bonaire card and getting this:
6,340.36 1,014.80 5.13 1,000.00 Binary Radio Pulsar Search (Arecibo, GPU) v1.39 (BRP4G-opencl-ati)

Magic is using an Nvidia 660Ti card and getting this:
28,610.53 8,929.42 51.11 3,333.00 Binary Radio Pulsar Search (Perseus Arm Survey) v1.39 (BRP5-cuda32-nv301)

I am using an Nvidia 760 card and getting this:
11,678.30 3,632.47 22.51 3,333.00 Binary Radio Pulsar Search (Perseus Arm Survey) v1.39 (BRP5-cuda32-nv301)

This all means I am doing units in half the time Magic is and getting the same credits. I am taking twice as long as tbret but getting three times the credit for it. I am buying the Nvidia 760 cards for about 200 US dollars apiece, which is similar to the links provided for the R9 cards. Here is a link to a 3 fan unit made by WindForce that is $225:

http://www.amazon.com/Gigabyte-GDDR5-2GB-WINDFORCE-GV-N760OC-2GD-REV2-0/dp/B00DGM8B6O/ref=sr_1_10?m=A2L77EE7U53NWQ&s=merchant&ie=UTF8&qid=1410953496&sr=1-10&keywords=nvidia+760

YES they are pre-owned cards, but not pre-used cards as in kept past the 14 day return for credit date. I have bought 3 or 4 of these pre-used cards now and have had no problems with any of them. I also have this specific model card and it works just fine. It is off line at the moment due to pc problems, but hopefully it will be back running again shortly.

I also do not own a pc with a pci-e(3) slot, my mb's are to old for that so I only have pci-e(2) slots. They are x16 slots but not the newer high speed slots and I do not have any problems, if there is a slowdown I wouldn't know it as I have nothing personally to compare it to.

David Rapalyea
David Rapalyea
Joined: 3 Jan 13
Posts: 79
Credit: 63886821
RAC: 0

Hi Majic Thanks for the

Hi Majic

Thanks for the info. My current rig is doing well enough but I have to tell you, getting four cards to run on one machine was a mind bender. Even my shop could not do it. Changing over to another brand of card is not in the cards these days. And the rigs seem efficient. About 150,000 stomes per day at 370 Watts.

Using my old mile stone unit thats 10,000 per day at about 25 watts.The four card machine might even get into the high teens but I will need to have a stable RAC before a final assesment. Ive only been up and running again for a few weeks and the RACs are still climing.

Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go

tbret
tbret
Joined: 12 Mar 05
Posts: 2115
Credit: 4862880478
RAC: 139333

RE: Maybe and maybe

Quote:


Maybe and maybe not...
tbret is using his AMD Bonaire card and getting this:
6,340.36 1,014.80 5.13 1,000.00 Binary Radio Pulsar Search (Arecibo, GPU) v1.39 (BRP4G-opencl-ati)

Magic is using an Nvidia 660Ti card and getting this:
28,610.53 8,929.42 51.11 3,333.00 Binary Radio Pulsar Search (Perseus Arm Survey) v1.39 (BRP5-cuda32-nv301)

I am using an Nvidia 760 card and getting this:
11,678.30 3,632.47 22.51 3,333.00 Binary Radio Pulsar Search (Perseus Arm Survey) v1.39 (BRP5-cuda32-nv301)

This all means I am doing units in half the time Magic is and getting the same credits. I am taking twice as long as tbret but getting three times the credit for it.

Well, maybe...

My times are two-at-a-time and I think I remember that Magic is running three at a time or maybe even four.

Are you doing two-at-a-time BRP5s on the 760?

EDIT: PS - I wondered how what you were comparing could possibly be the case...

My reference was to my i5 machine with the Pitcairn in it. It is also doing two-at-a-time and the run is more like 3,300 seconds, meaning each is taking 1,650 seconds (or so; I'm generalizing). That's 2,000 credits every 3,300 seconds.

The Bonaire is a R7 260X and cost $119. It's also in a system I would considered crippled by its x1 PCIe slot, slow DDR2 RAM, and a weak CPU.

mikey
mikey
Joined: 22 Jan 05
Posts: 12680
Credit: 1839083724
RAC: 3916

RE: RE: Maybe and maybe

Quote:
Quote:


Maybe and maybe not...
tbret is using his AMD Bonaire card and getting this:
6,340.36 1,014.80 5.13 1,000.00 Binary Radio Pulsar Search (Arecibo, GPU) v1.39 (BRP4G-opencl-ati)

Magic is using an Nvidia 660Ti card and getting this:
28,610.53 8,929.42 51.11 3,333.00 Binary Radio Pulsar Search (Perseus Arm Survey) v1.39 (BRP5-cuda32-nv301)

I am using an Nvidia 760 card and getting this:
11,678.30 3,632.47 22.51 3,333.00 Binary Radio Pulsar Search (Perseus Arm Survey) v1.39 (BRP5-cuda32-nv301)

This all means I am doing units in half the time Magic is and getting the same credits. I am taking twice as long as tbret but getting three times the credit for it.

Well, maybe...

My times are two-at-a-time and I think I remember that Magic is running three at a time or maybe even four.

Are you doing two-at-a-time BRP5s on the 760?

EDIT: PS - I wondered how what you were comparing could possibly be the case...

My reference was to my i5 machine with the Pitcairn in it. It is also doing two-at-a-time and the run is more like 3,300 seconds, meaning each is taking 1,650 seconds (or so; I'm generalizing). That's 2,000 credits every 3,300 seconds.

The Bonaire is a R7 260X and cost $119. It's also in a system I would considered crippled by its x1 PCIe slot, slow DDR2 RAM, and a weak CPU.

Yes I am running 2 at a time to get those numbers. And yes my 760 here is in an AMD 6 core machine with 16gb of ram. That is as good as any of my machines get though, I do have a couple others just like it or very similar, and then I too go down from there.

MAGIC you have mail!

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.