From: "Axel Huebl a.huebl@hzdr.de" Received: from [149.220.60.110] (account huebl@hzdr.de [149.220.60.110] verified) by hzdr.de (CommuniGate Pro SMTP 6.2.7i) with ESMTPSA id 20749757 for picongpu-users@hzdr.de; Tue, 04 Sep 2018 10:42:02 +0200 Subject: Re: PIConGPU performance across different GPUs To: picongpu-users@hzdr.de References: Organization: HZDR Message-ID: Date: Tue, 4 Sep 2018 10:42:26 +0200 MIME-Version: 1.0 In-Reply-To: Content-Type: text/plain; charset=utf-8 Content-Language: en-US Content-Transfer-Encoding: 8bit Yep, during the last generations (k20 - k80, p100, v100) we got computation-wise always a speedup one would expect, proportional to the theoretical flop/s of the different models. Regarding your off-list question on Hypnos K20: they only have 5 GByte and prior to P100 one also needs to take into account that activated ECC RAM costs you around 12%. Hypnos has 64 K20 cards and 36 K80 cards (72 "GPUs"). Both queues have four cards per node, respectively (16 nodes in the "k20" queue and 9 nodes in the "k80" queue). https://www.hzdr.de/db/Cms?pOid=29813&pNid=852 Credit: one node of the nine k80 nodes we are hosting with and for our collaborators at HI Jena / GSI. In return, they can also use the other k80 nodes. I can imagine similar concepts for future queues, e.g. we are building a new cluster that has P100 nodes and adding additional nodes from collaborators for shared usage would be cool to allow larger simulations. Best, Axel On 9/3/18 12:45 PM, Andrei Berceanu berceanu@runbox.com wrote: > Thank you René. > So are the 64 K80 boards on Hypnos 64 cards (with 2 GPUs each), or 64 > GPUs (32 actual cards)? > > Best, > Andrei > > On Mon, Sep 3, 2018 at 11:31 AM René Widera r.widera@hzdr.de > > wrote: > > Hi Andrei, > > I have currently no performance data for the P100 available but if I > remember correct one P100 is two times faster than one k80 (with on K80 > I mean one card with two sockets). > For V100 I have currently no numbers in my head but due to improvements > of the shared memory atomics a V100 should be faster than a P100. > > best regards, > > René > > On 08/31/2018 09:52 AM, Andrei Berceanu berceanu@runbox.com > wrote: > > Dear all, > > > > In you experience, how does PIConGPU performance vary across > different > > Tesla cards, such as the K80, P100 and V100? > > > > Is it fair to say, for instance, that one P100 is equivalent to > two K80s? > > > > Best regards, > > Andrei > > -- > René Widera > Phone: +49 351 260 3543 > Institute of Radiation Physics > http://www.hzdr.de/crp > Helmholtz-Zentrum Dresden - Rossendorf (HZDR) > Bautzner Landstr. 400 | 01328 Dresden | Germany > Board of Directors: Prof. Dr. Dr. h. c. Roland Sauerbrey, Dr. Ulrich > Breuer > Company Registration Number VR 1693, Amtsgericht Dresden > > ############################################################# > This message is sent to you because you are subscribed to >   the mailing list >. > To unsubscribe, E-mail to: > > To switch to the DIGEST mode, E-mail to > > > To switch to the INDEX mode, E-mail to > > Send administrative queries to  > > > -- Axel Huebl Phone: +49 351 260 3582 Institute of Radiation Physics http://www.hzdr.de/crp Helmholtz-Zentrum Dresden - Rossendorf (HZDR) Bautzner Landstr. 400 | 01328 Dresden | Germany Board of Directors: Prof. Dr. Dr. h. c. Roland Sauerbrey, Dr. Ulrich Breuer Company Registration Number VR 1693, Amtsgericht Dresden