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


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.

AuthorMessage
Profile gabefair
Send message
Joined: Dec 30 08
Posts: 2
Credit: 3,857,921
RAC: 1,855
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
    Avatar
    Send message
    Joined: Jan 18 05
    Posts: 486
    Credit: 84,278,090
    RAC: 179,519
    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: Oct 3 11
      Posts: 205
      Credit: 69,388,808
      RAC: 94,427
      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: Nov 20 12
        Posts: 175
        Credit: 169,662,558
        RAC: 103,692
        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: Jun 10 06
          Posts: 48
          Credit: 5,048,575
          RAC: 7,699
          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: Jun 14 12
            Posts: 70
            Credit: 142,553,774
            RAC: 390,944
            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: Jan 22 05
              Posts: 960
              Credit: 20,668,820
              RAC: 190,465
              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
                Forum moderator
                Avatar
                Send message
                Joined: Dec 1 05
                Posts: 3320
                Credit: 27,254,204
                RAC: 21,924
                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

                  Alec
                  Send message
                  Joined: Dec 5 10
                  Posts: 54
                  Credit: 41,007,166
                  RAC: 10,385
                  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: Sep 29 09
                    Posts: 30
                    Credit: 77,234,473
                    RAC: 3,846
                    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: Jan 14 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: Feb 21 05
                        Posts: 15
                        Credit: 2,027,858
                        RAC: 50
                        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!
                          ____________

                          Alec
                          Send message
                          Joined: Dec 5 10
                          Posts: 54
                          Credit: 41,007,166
                          RAC: 10,385
                          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: Feb 28 05
                            Posts: 117
                            Credit: 108,110,132
                            RAC: 363,957
                            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: Nov 20 12
                              Posts: 175
                              Credit: 169,662,558
                              RAC: 103,692
                              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: Feb 28 05
                                Posts: 117
                                Credit: 108,110,132
                                RAC: 363,957
                                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: Sep 21 10
                                  Posts: 34
                                  Credit: 387,115,970
                                  RAC: 955,894
                                  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: Mar 1 05
                                    Posts: 390
                                    Credit: 24,459,474
                                    RAC: 39,477
                                    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: Jan 5 13
                                      Posts: 1
                                      Credit: 479,011
                                      RAC: 92
                                      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: Feb 28 05
                                        Posts: 117
                                        Credit: 108,110,132
                                        RAC: 363,957
                                        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.

                                          Profile Alex
                                          Send message
                                          Joined: Mar 1 05
                                          Posts: 390
                                          Credit: 24,459,474
                                          RAC: 39,477
                                          Message 123899 - Posted 7 Apr 2013 22:11:38 UTC - in response to Message 123896.

                                            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.


                                            This is how I understood it. The runtime for his wu's are reported as ~ 3000sec which points more to my calculation than to 30 wu's/h
                                            ____________

                                            Post to thread

                                            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 © 2014 Bruce Allen