 |
 |
You are completely right. I was checking that in the specifications of the machine. I think that they provide the fifth slot because the typical user of the system would have 4 Teslas in the x16 slots and a normal card in the x8. x8 should be more than enough for everyday tasks :). On 18/07/16 15:05, Ludtke, Steven J wrote: > Not to butt in here, but unless I'm mistaken, there aren't any standard chipsets that can provide 5 channels of PCIe x16. Motherboards which offer a 5th x16 slot actually support only x8 transfers, meaning transfers to/from GPU RAM will be 1/2 as fast on this card. Whether this has an impact will depend on the implementation. > >> On Jul 18, 2016, at 8:56 AM, Sjors Scheres <[log in to unmask]> wrote: >> >> ***CAUTION:*** This email is not from a BCM Source. Only click links or open attachments you know are safe. >> >> ---------------------------------------------------------------------- >> If you already can fit 5, you might want to run your relion jobs with 5 as well. >> :-) >> S >> >> On 07/18/2016 12:50 PM, Alberto Riera wrote: >>> Hi Sjors, >>> >>> Thanks for the advice! Luckily, we can use the 4 cards for Relion. However, since this system actually allows to fit an extra GPU I was considering if it would be worth it :) >>> >>> Alberto >>> >>> >>> On 18/07/16 12:41, Sjors Scheres wrote: >>>> Hi Alberto, >>>> Odd number of cards will be used equally efficiency as even numbers for 2D and 3D classification. It's only the 3D auto-refine that makes better use of even number of cards (as each half-set is processed by a different MPI process). If you need 1 card for something else, I would probably choose to run relion with 3 cards rather than complicating your setup with 5 cards.... >>>> HTH, >>>> Sjors >>>> >>>> On 07/18/2016 10:42 AM, Alberto Riera wrote: >>>>> Thank you for all the advice! >>>>> >>>>> I was thinking a little bit more about this computer and I wonder if it would be advantageous for us to install an additional GPU. I know that Relion uses the GPUs better in pairs, so the idea would be to let 4 cards just for it and use the spare one for some other tasks if needed. Alternatively, we could use the fifth card with Relion as well, although it would be less efficient than having six of them. >>>>> >>>>> >>>>> On 14/07/16 16:01, Craig Yoshioka wrote: >>>>>> Hi Alberto, >>>>>> >>>>>> I’d like to add, it might be worthwhile to keep a few of those CPUs and RAM on the side for ZFS. I’d highly recommend turning on ZFS LZ4 compression, especially on raw frame stacks it’ll save a lot of space and be faster I/O. >>>>>> >>>>>> >>>>>> >>>>>>> On Jul 14, 2016, at 5:09 AM, Erik Lindahl <[log in to unmask] <mailto:[log in to unmask]>> wrote: >>>>>>> >>>>>>> Hi Alberto, >>>>>>> >>>>>>> On vacation, so sorry if I'm a bit brief. >>>>>>> >>>>>>> I think your machine will work great if it's intended as a staging server where you do all/most processing. We have a highly similar one, where we also got an external 44-disk enclosure for storage. >>>>>>> >>>>>>> However, if you want to get more than one machine, we prefer to separate CPU-heavy and GPU-heavy nodes. The reason is that even the GPU code will use CPUs, so you will hurt your GPU performance a lot if you start to run a CPU job on the same machine - and if you don't, you waste those expensive CPUs. For this reason we bought one set of CPU-only nodes with fast processors and then cheap CPUs for the GPU nodes - on the whole I think that provides better value for cluster nodes. >>>>>>> >>>>>>> We no longer use RAID cards, but have moved to ZFS instead. If you use ubuntu 16.04 it even comes built-in. >>>>>>> >>>>>>> The released version of oculus rift unfortunately no longer supports plain display mode, so unless a program has been adapted to their devkit, you won't be able to use it. >>>>>>> >>>>>>> Cheers, >>>>>>> >>>>>>> Erik >>>>>>> >>>>>>> >>>>>>> >>>>>>> Erik Lindahl <[log in to unmask] <mailto:[log in to unmask]>> >>>>>>> Professor of Biophysics >>>>>>> Science for Life Laboratory >>>>>>> Stockholm University & KTH >>>>>>> Office (SciLifeLab): +46 8 524 81567 >>>>>>> Cell (Sweden): +46 73 4618050 >>>>>>> Cell (US): +1 267 3078746 >>>>>>> >>>>>>> >>>>>>> On 14 juli 2016, at 07:24, Alberto Riera <[log in to unmask] <mailto:[log in to unmask]>> wrote: >>>>>>> >>>>>>>> Hi guys! >>>>>>>> >>>>>>>> After a lot of days working on this we finally have defined a system to work with the GPU implementation of Relion and it would be great it you could tell us your opinion. Currently it is a bit on the overkill side from the CPU point of view, but we managed to obtain a very good price and, on the other hand, it is future proof in this regard and in case we need to install more RAM (it goes up to 2TB). >>>>>>>> >>>>>>>> So, it is going to be a Supermicro 4U GPU chassis (workstation format) with the following: >>>>>>>> >>>>>>>> 2x Intel E5-2697v4 2.3Ghz Processors (36 cores) >>>>>>>> 8x Supermicro 16GB DDR4 ECC REG Memory >>>>>>>> 1x 960GB SATA Read Intensive Enterprise Level SSD >>>>>>>> 4x 2TB SATA Mixed Workload Enterprise Level SSD (RAID 5) >>>>>>>> 1x LSI 9361-8i RAID Controller inc CacheVault >>>>>>>> 4x NVidia GTX 1080 >>>>>>>> 1x Mellanox Single Port 10Gb SFP+ NIC >>>>>>>> >>>>>>>> Any opinions on it? >>>>>>>> >>>>>>>> On the other hand, we are looking to do some 3D visualization with the system and I wonder if any of you has experience with Linux and the new VR headsets. They seem very promising but I still did not have the chance to test them. If you have any other preferred option, suggestions are very welcome :). >>>>>>>> >>>>>>>> Thanks! >>>>>>>> >>>>>>>> Alberto >>>>>>>> >>>>>>>> -- >>>>>>>> Alberto Riera PhD >>>>>>>> Imperial College London >>>>>>>> Faculty of Medicine >>>>>>>> Institute of Clinical Sciences (CRB room 3006) >>>>>>>> Hammersmith Hospital Campus >>>>>>>> Du Cane Road >>>>>>>> London W12 0NN >>>>>>>> >>>>>>>> [log in to unmask] >> -- >> Sjors Scheres >> MRC Laboratory of Molecular Biology >> Francis Crick Avenue, Cambridge Biomedical Campus >> Cambridge CB2 0QH, U.K. >> tel: +44 (0)1223 267061 >> https://urldefense.proofpoint.com/v2/url?u=http-3A__www2.mrc-2Dlmb.cam.ac.uk_groups_scheres&d=DQIDaQ&c=ZQs-KZ8oxEw0p81sqgiaRA&r=GWA2IF6nkq8sZMXHpp1Xpg&m=MC89isJA_EGf3djZJthITxDNnJee1YzfW-XB2mSt-Jo&s=Xis_VLjzOTxIkh7HkWOiy4FrKog2KOi7UX0yxEmb4KA&e= > ---------------------------------------------------------------------------- > Steven Ludtke, Ph.D. > Professor, Dept of Biochemistry and Mol. Biol. (www.bcm.edu/biochem) > Co-Director National Center For Macromolecular Imaging (ncmi.bcm.edu) > Co-Director CIBR Center (www.bcm.edu/research/cibr) > Baylor College of Medicine > [log in to unmask] > > > > > -- Alberto Riera PhD Imperial College London Faculty of Medicine Institute of Clinical Sciences (CRB room 3006) Hammersmith Hospital Campus Du Cane Road London W12 0NN [log in to unmask]