Can Einstein handle 3.5 PFlops offered by NVIDIA's K20? Cuz I'm thinking about buying one.

log in

Advanced search

Message boards : Cruncher's Corner : Can Einstein handle 3.5 PFlops offered by NVIDIA's K20? Cuz I'm thinking about buying one.

1 · 2 · Next
Author Message
Profile gabefair
Send message
Joined: 30 Dec 08
Posts: 2
Credit: 10,652,536
RAC: 6,969
Message 123524 - Posted: 19 Mar 2013, 11:56:28 UTC
Last modified: 19 Mar 2013, 12:06:24 UTC

NVIDIA's K20 GPU has 2496 CUDA cores and offers a Peak single precision floating point performance (board): 3.52 Tflops for only 225 watts!

My question is, Can Einstein use all of those cores if I were to buy one? Would this GPU be used fully?

What would happen to the Einstein@Home community if one or two of the GPUs were to come online?

Profile MAGIC Quantum Mechanic
Avatar
Send message
Joined: 18 Jan 05
Posts: 1057
Credit: 273,368,774
RAC: 257,217
Message 123531 - Posted: 19 Mar 2013, 13:46:34 UTC - in response to Message 123524.

Well I wouldn't mind testing one here as long as it has MB with 16GB ram and the fastest multi-core CPU I could have at the same time.

It would cost about the same as buying 11 nVidia GeForce GTX 660Ti 2GB SC

And I have found out that the cards I have act different (slower) if I try to run BRP X2 along with 4 GRP's (or 3 with a 3-core)

My 650Ti SC is running with one of my older 3-core and it does ok until I tried to run the 3GRP's but I have it;s twin running with a 550TI OC and it was still doing ok running the GRP's at the same time (no free cores)

The 660Ti runs ok that way too.

Funny thing is the laptop I am on right now runs BRP X2 cudas and 8 GRP's at the same time 24/7 for almost a year now and it runs just fine without a free core.

So like I said I get tempted to test cards like that Tesla K20 but I wouldn't be surprised if it doesn't do what you would hope for when you pay $3,499

Horacio
Send message
Joined: 3 Oct 11
Posts: 205
Credit: 80,557,243
RAC: 0
Message 123533 - Posted: 19 Mar 2013, 15:03:30 UTC - in response to Message 123524.

I cant know in advance if that GPU is worth for number crunching, but for sure it wont affect the project in any way...
The project was able to handle the massive power of lots of GPUs some time ago when SETI was offline, so just one extra GPU is not going to be noticed, no matter how fast is.

But, if you are considering that GPU just for number crunching in Einstein, then I fear you will be dissapointed... Teslas have a really high cost because their speed on double pressicion which is not needed on Einstein.

____________

Neil Newell
Send message
Joined: 20 Nov 12
Posts: 176
Credit: 169,699,457
RAC: 0
Message 123537 - Posted: 19 Mar 2013, 15:32:49 UTC - in response to Message 123533.
Last modified: 19 Mar 2013, 15:33:20 UTC

...plus they are engineered for reliability over outright speed, so clocked slower and use ECC RAM.

Probably better to spend the money on GTX Titan - less than 1/3 the price, and 2688 cores.

Profile microchip
Send message
Joined: 10 Jun 06
Posts: 48
Credit: 6,947,269
RAC: 0
Message 123538 - Posted: 19 Mar 2013, 16:04:59 UTC

Yeah, I also recommend the Titan over a Tesla. Over at GPUGRID they're working on an application that will run optimal on the Titan. No idea if this is needed for Einstein too (my guess it isn't)
____________

Team Belgium

mountkidd
Send message
Joined: 14 Jun 12
Posts: 104
Credit: 637,584,425
RAC: 317,783
Message 123540 - Posted: 19 Mar 2013, 18:14:59 UTC

For E@H GPU processing, buswidth and bandwidth appear to have a much greater impact than core count. GTX 670/680 are close in bandwidth, GFlops/Watt to the K20 and almost an order of magnitude lower in cost. If raw power is what you are after consider the HD7970 as its bandwidth is ~30% higher than the 3 cards mentioned above and is spec'd at 3.7 TFlops.

Gord

mikey
Avatar
Send message
Joined: 22 Jan 05
Posts: 2575
Credit: 160,594,730
RAC: 124,894
Message 123560 - Posted: 20 Mar 2013, 13:14:31 UTC - in response to Message 123540.

For E@H GPU processing, buswidth and bandwidth appear to have a much greater impact than core count. GTX 670/680 are close in bandwidth, GFlops/Watt to the K20 and almost an order of magnitude lower in cost. If raw power is what you are after consider the HD7970 as its bandwidth is ~30% higher than the 3 cards mentioned above and is spec'd at 3.7 TFlops.

Gord


I can't say about Einstein but I CAN say the AMD 7970 gpu is SCREAMINGLY fast!! Comparing it to the 5770 it replaced I went from doing units in about 1.5 hours to 10 MINUTES!! I am using it at DistRTgen so I do not even know if it CAN work here, there are some projects that just haven't made it work for them yet.
Profile Mike Hewson
Volunteer moderator
Avatar
Send message
Joined: 1 Dec 05
Posts: 5081
Credit: 41,740,858
RAC: 9,771
Message 123572 - Posted: 20 Mar 2013, 23:41:42 UTC - in response to Message 123524.
Last modified: 21 Mar 2013, 3:10:07 UTC

What would happen to the Einstein@Home community if one or two of the GPUs were to come online?

We'd be very happy !

Never knock back a petaflop, that's what I always say ... :-) :-)

IMHO the Titan specs are awesome, especially as Neil says on the reliability side but also power use. Naturally it's performance is subject to moderation by the system it's in. Having said that though it is a major outlay for anybody, definitely a bleeding edge item. It'll be the most expensive single item within the case for sure.

[musings]
I think we're moving/merging toward an era of 'massively parallel coprocessors, some of which also do graphics'. By that I mean I'm not aware of an application that would purely use such a card to the full extent of it graphics performance alone. Could well be my ignorance, but many FPS can you justify for some video game or 3D visualisation ? Does anyone know of an example ? This is relevant in that for such an outlay, is there a business ( non-supercomputing ) purpose that justifies the purchase per se, and can maybe do E@H et al on the side .....
[/musings]

Cheers, Mike.

( edit ) Wow, they're talking 3-way SLI here.
____________
"I have made this letter longer than usual, because I lack the time to make it short." - Blaise Pascal
Gamboleer
Send message
Joined: 5 Dec 10
Posts: 173
Credit: 115,027,106
RAC: 460,704
Message 123620 - Posted: 24 Mar 2013, 1:21:40 UTC - in response to Message 123524.

<<What would happen to the Einstein@Home community if one or two of the GPUs were to come online?>>

Fire and brimstone coming down from the skies! Rivers and seas boiling! Forty years of darkness! Earthquakes! Volcanoes! The dead rising from the grave! Human sacrifice! Dogs and cats living together! Mass hysteria!

HenkM
Send message
Joined: 29 Sep 09
Posts: 32
Credit: 165,023,169
RAC: 0
Message 123621 - Posted: 24 Mar 2013, 1:41:59 UTC

If you have a look here
http://einstein.phys.uwm.edu/results.php?hostid=5728548&offset=0&show_names=1&state=3&appid=0
choosing between AMD/ATI 7970 and Nvidea Titan is easy in my opinion.
Here in NL a 7970 costs just halve of a Titan.

Profile Nobody316
Avatar
Send message
Joined: 14 Jan 13
Posts: 141
Credit: 2,008,126
RAC: 0
Message 123622 - Posted: 24 Mar 2013, 1:52:33 UTC

http://www.newegg.com/Product/Product.aspx?Item=N82E16814131483&IsVirtualParent=1

PowerColor AX7990 6GBD5-2DHJ Radeon HD 7990 6GB 384-bit x2 GDDR5 PCI Express 3.0 x16 HDCP Ready CrossFireX Support Video Card looks quite nice.

Revolutionize your gaming experience with the PowerColor AX7990 VGA card. Powered by dual class-leading AMD Radeon HD 7970 engines, the PowerColor AX7990 tackles your most demanding HD titles with monstrous performance from 2048 x 2 Stream Processors and total of 6GB 384-bit x 2 GDDR5 memory. Your games cease to be the same with redefined realism and skip-free experience even at the highest setting. Dual BIOS design enables you to push the performance potential to limit with a press of one button. An array of AMD graphics technology such as AMD Eyefinity, AMD HD 3D and AMD PowerTune offers you more than one way to elevate your gaming experience.


The price is not too bad either but it looks like it would be cheaper just to buy 2 of the 7970 cards unless ya buy 2 of the 7990 which would = 4 of the 7970's. I agree the 7970 seems quite fast and will be something I will be checking out in the near future.
____________
PC setup MSI-970A-G46 AMD FX-8350 8 core OC'd 4.45GHz 16GB ram PC3-10700 Geforce GTX 650Ti Windows 7 x64 Einstein@Home
Ivailo Bonev
Send message
Joined: 21 Feb 05
Posts: 19
Credit: 2,847,526
RAC: 2
Message 123633 - Posted: 24 Mar 2013, 23:22:17 UTC - in response to Message 123620.

<<What would happen to the Einstein@Home community if one or two of the GPUs were to come online?>>

Fire and brimstone coming down from the skies! Rivers and seas boiling! Forty years of darkness! Earthquakes! Volcanoes! The dead rising from the grave! Human sacrifice! Dogs and cats living together! Mass hysteria!


I had good laugh on this! :D Thank you!
____________
Gamboleer
Send message
Joined: 5 Dec 10
Posts: 173
Credit: 115,027,106
RAC: 460,704
Message 123649 - Posted: 26 Mar 2013, 2:43:00 UTC - in response to Message 123633.

<<I had good laugh on this! :D Thank you!>>

It was funnier in Ghostbusters :)

Profile Beyond
Avatar
Send message
Joined: 28 Feb 05
Posts: 117
Credit: 148,719,734
RAC: 59
Message 123738 - Posted: 1 Apr 2013, 14:59:25 UTC - in response to Message 123524.

NVIDIA's K20 GPU has 2496 CUDA cores and offers a Peak single precision floating point performance (board): 3.52 Tflops for only 225 watts!
My question is, Can Einstein use all of those cores if I were to buy one? Would this GPU be used fully?
What would happen to the Einstein@Home community if one or two of the GPUs were to come online?

Titan should be faster for DC at 1/3 the cost (but still too expensive). 2 x HD 79xx should be faster than either at even less $$$. Power efficiency should be no better than the lower cost alternatives IMO.
Neil Newell
Send message
Joined: 20 Nov 12
Posts: 176
Credit: 169,699,457
RAC: 0
Message 123791 - Posted: 4 Apr 2013, 7:45:52 UTC - in response to Message 123738.

I see there are hosts sporting Titans now - and here's a host with two of them.

Profile Beyond
Avatar
Send message
Joined: 28 Feb 05
Posts: 117
Credit: 148,719,734
RAC: 59
Message 123795 - Posted: 4 Apr 2013, 14:39:31 UTC - in response to Message 123791.

I see there are hosts sporting Titans now - and here's a host with two of them.

It would be very interesting to know how many WUs they're running concurrently, in fact it would be nice to have this information in the results file. If it's there I'm missing it...
Gavin
Send message
Joined: 21 Sep 10
Posts: 117
Credit: 2,502,208,440
RAC: 3,801,762
Message 123798 - Posted: 4 Apr 2013, 16:54:21 UTC
Last modified: 4 Apr 2013, 16:56:55 UTC

Quick look at those hosts last contact log's suggests 5 at a time per card.
GPU utilisation is set at 0.2 with an effective runtime just shy of 10 mins. per task.

Profile Alex
Send message
Joined: 1 Mar 05
Posts: 438
Credit: 47,775,980
RAC: 1,256
Message 123800 - Posted: 4 Apr 2013, 17:27:17 UTC - in response to Message 123798.

Quick look at those hosts last contact log's suggests 5 at a time per card.
GPU utilisation is set at 0.2 with an effective runtime just shy of 10 mins. per task.


Which means 6 wu's per hour.

My Sapphire HD7870 XT finishes 4 wu's per hour and I paid ~216.- Euro.
Power consumption < 180W when running 2 wu's at a time

Supporting CPU: i3 ivy bridge @3.3GHz, slot: pci-e3 @ x8
http://einstein.phys.uwm.edu/show_host_detail.php?hostid=6633241

A look into the statistics / best computers gives a clear picture:
places 1-5 AMD, place 6 a PC with 4 Teslas C2075.

I'm happy with my decision.
____________
Jesper Sorensen
Send message
Joined: 5 Jan 13
Posts: 1
Credit: 3,787,458
RAC: 1,129
Message 123895 - Posted: 7 Apr 2013, 21:19:41 UTC - in response to Message 123800.
Last modified: 7 Apr 2013, 21:20:42 UTC

5 WUs at one time, ~10 mins/WU equals 30 WUs/hour .....

Profile Beyond
Avatar
Send message
Joined: 28 Feb 05
Posts: 117
Credit: 148,719,734
RAC: 59
Message 123896 - Posted: 7 Apr 2013, 21:23:20 UTC - in response to Message 123895.

5 WUs at one time, ~10 mins/WU equals 30 WUs/hour .....

Maybe not so clear, it's 50 minutes or so for 5 tasks: 6 tasks/hour.
1 · 2 · Next

Message boards : Cruncher's Corner : Can Einstein handle 3.5 PFlops offered by NVIDIA's K20? Cuz I'm thinking about buying one.


Home · Your account · Message boards

This material is based upon work supported by the National Science Foundation (NSF) under Grants PHY-1104902, PHY-1104617 and PHY-1105572 and by the Max Planck Gesellschaft (MPG). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the investigators and do not necessarily reflect the views of the NSF or the MPG.

Copyright © 2016 Bruce Allen