CUDA and openCL Benchmarks

MAGIC Quantum Mechanic
MAGIC Quantum M...
Joined: 18 Jan 05
Posts: 1907
Credit: 1438686145
RAC: 1226209

Hello again Dskag, I have


Hello again Dskag,

I have had my NVIDIA GeForce GTX 650 Ti (2047MB) driver: 31407 running for a while now so I finally got it to work as fast as I can.

Running BRP cuda X2 and it is just on a 3-core host that at the same time uses 2-cores +virtual machine for T4T

So using the EVGA Precision X I got the average at 3,175 seconds as long as I am not on that host doing something else at the same time.

(just over 53mins BRP X2)

http://einsteinathome.org/host/6661572

* MAGIC

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5874
Credit: 118251749696
RAC: 24783303

RE: (just over 53mins BRP

Quote:
(just over 53mins BRP X2)


650Ti's can do a bit better than that. This results list shows one doing BRP4 tasks (x2), in very close to 49 minutes (~2950 secs). The host is a 3570K running at 4.2GHz with all CPU cores running FGRP2 tasks (no free cores).

My experience with this setup is that there is no further gain in GPU production if one or more cores are stopped from running CPU tasks. I also find that even fewer and lower powered CPU cores (like a dual core G645 processor for example) can maintain the same GPU performance. Of course, power consumption rises and power efficiency decreases by crunching CPU tasks but, at least until a GPU app is available for FGRP2 tasks, I'll continue running the CPU cores full bore. To me, it's important to continue supporting the CPU only apps here.

Cheers,
Gary.

mountkidd
mountkidd
Joined: 14 Jun 12
Posts: 177
Credit: 12720752973
RAC: 4680101

RE: I have had my NVIDIA

Quote:

I have had my NVIDIA GeForce GTX 650 Ti (2047MB) driver: 31407 running for a while now so I finally got it to work as fast as I can.

I have been running my 550/670 cards with the 31422/31407/31090 drivers and have found that 31090 produces slightly better results. If you are up to some tweaking, this might be something to look at.

Gord

Beyond
Beyond
Joined: 28 Feb 05
Posts: 123
Credit: 2476987679
RAC: 4811529

RE: RE: (just over 53mins

Quote:
Quote:
(just over 53mins BRP X2)

650Ti's can do a bit better than that. This results list shows one doing BRP4 tasks (x2), in very close to 49 minutes (~2950 secs). ... with all CPU cores running FGRP2 tasks (no free cores).


Could XP vs. Linux be an explanation?

Sunny129
Sunny129
Joined: 5 Dec 05
Posts: 162
Credit: 160342159
RAC: 0

i'm also wonderng if there

i'm also wonderng if there aren't other factors contributing to Gary's situation. While he seems to have found that freeing up some additional CPU cores does nothing to improve the crunching efficiency of his 650 Ti, several of us have found the exact opposite to be quite true - that freeing up additional CPU cores can in fact help improve GPU crunching efficiency. so i'm wondering if that doesn't have more to do with his platform specs, such as Windows vs Linux, actual CPU throughput, chipset X vs chipset Y, PCIe 2.0 vs PCIe 3.0, etc.

mountkidd
mountkidd
Joined: 14 Jun 12
Posts: 177
Credit: 12720752973
RAC: 4680101

RE: Could XP vs. Linux be

Quote:
Could XP vs. Linux be an explanation?


In part, but the underlying feeder hardware could also be a major contributor. The difference in BRP elapsed time is only 7%. Looking at the benchmark numbers for MAGIC: 2756 FP/s & 6008 Int/s vs Gary's: 3889 FP/s & 18673 Int/s there is a huge performance disparity - 1.4x & 3x respectively . Using 2 free cores might help close the gap and would certainly be easy to test.

Gary... I have several i5-3570K that are not overclocked and benchmark 3665 FP/s and 12658 Int/s. Your OC is ~20% higher but your Int number is close to 50% higher. Are you doing anything else to get this increase?

Gord

MAGIC Quantum Mechanic
MAGIC Quantum M...
Joined: 18 Jan 05
Posts: 1907
Credit: 1438686145
RAC: 1226209

RE: RE: I have had my

Quote:
Quote:

I have had my NVIDIA GeForce GTX 650 Ti (2047MB) driver: 31407 running for a while now so I finally got it to work as fast as I can.

I have been running my 550/670 cards with the 31422/31407/31090 drivers and have found that 31090 produces slightly better results. If you are up to some tweaking, this might be something to look at.

Gord

I think I will have to give that a try with the 650Ti first and I have a couple 550Ti's too and I see you have a couple with that driver so maybe those will be next and I also have a 660Ti.

So far both of my 550Tis have been faster than the 650Ti and they are all OC'd cards.

Thanks Gord,

*Magic*

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5874
Credit: 118251749696
RAC: 24783303

RE: Gary... I have several

Quote:

Gary... I have several i5-3570K that are not overclocked and benchmark 3665 FP/s and 12658 Int/s. Your OC is ~20% higher but your Int number is close to 50% higher. Are you doing anything else to get this increase?

Gord


Gord,

Apart from overclocking, I've made no other tweaks to anything. Before I'd even bought the CPU, I had regularly seen on overclocking sites that it would do 4.2GHz easily and probably 4.5GHz or better with a bit of extra effort. So, when I first fired it up it was at 4.2GHz and it has never been at stock. I didn't even bother stress testing it but simply put it straight to crunching, figuring that I'd soon know if it had a problem. I always intended to revisit for some extra tweaking but have never found the time to do it. There's probably a bit more performance to be had but I'm very happy with the way it is. Apart from a BOINC bug with version 6.12.34, the machine has been quite stable.

I would guess that the benchmarks inconsistencies are a function of the OS difference. Some years ago, when I did run a lot of XP machines, there were always discrepancies on the same hardware between WinXP and Linux. Some of that may have been due to different BOINC versions as well.

With regard to getting better GPU performance by having free CPU core(s), I should emphasise that my comments are in relation to GTX650 and 650Ti GPUs when running BRP4 tasks 2x. I also have a HD7770 running 2x that does suffer if there isn't a free core. I would guess that higher end GPUs would also benefit even more significantly.

I've been quite surprised to see how well the 650 does on certain older motherboards too. For comparison, my best performance has been around 59 mins (2x) when running on a H61 board with a Sandy Bridge generation CPU (PCIe 2.0). I recently put one in a much older PCIe 1.x system powered by a 4 year old Sempron processor - single core but unlocked through the BIOS to a dual core. Once again, both CPU cores are crunching but the GPU is still performing very close to the best I get on a modern system. The results are here.

Cheers,
Gary.

mountkidd
mountkidd
Joined: 14 Jun 12
Posts: 177
Credit: 12720752973
RAC: 4680101

I recently upgraded from

I recently upgraded from Boinc 7.0.28 to 7.0.64 and was very surprised by the drop in elapsed time for BRP tasks. After running for a few days, statistics were collected, summarized and are presented below:

Elapsed and cpu times are an average of 50 credited readings.

I am trying to understand why there should be such an improvement in performance and at this point can only speculate. Boinc is outside the BRP executables which is where I would expect to see performance gains. The BRP...exe is date stamped Feb 1/13 so there haven't been any recent changes there. Perhaps the newer Boinc is now enabling something in the BRP executable.

In any case, I am very pleased with the extra throughput being experienced and would recommend the 7.0.64 upgrade to those that haven't done so.

Gord

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

Hello Everyone, Beyond

Hello Everyone,

Beyond asked me to report my rig and production on this thread.

SUMMARY: Six BRP tasks in just under 75 minutes @ just less then 215 Watts

-- ASRock Seven Slot Gaming Motherboard
-- 32 pcie 3.0 lanes branched for two clusters [x16 x8 x16] [x8 x16 x1 x16]
-- Intel Ivy Bridge i3 E3220 @3.3 GHZ [two cores, two threads each]
-- Seasonic ATX 360 watt 80+Gold
-- Solid State Drive, 30 Gig
-- 4 gig ddr
-- Windows Home 7-64
-- Three cordless Asus GTX 650s GPU mounted in slots 1 and 3 and 7

I am not proficient in electronics, but took up Einstien because I think neutron stars are eally really neat. I keep asking if a neutron star can become a black hole by simple gradual gravitational mass accretion, or if such a colapse requires large mass colliding at relativistic speeds. I have not heard a convincing arguement either way.

However, I WAS on the proper side of Hawking Radiation. I agreed with Suskind before I ever heard of Suskind.

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

Comment viewing options

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