Dear All, It was suggested I post a message from yesterday here. My main concern is do the Radeons work? From the OpenCL site, they can but I have yet to find something that at least on paper looks as good as CUDA. In any case, I am trying by weeks end to at least have all the parts for a costum PC written down and priced. Any suggestions with the system (I'm looking at Lunix with OpenCL as of now), or other are appriciated. I did recieve a couple good pointers as of now, and am into Molecular dynamics (along with any wet labs whom might be interested in hiring :>) , as I am a structural and biophysics based person. OG Post: Dear ALL, I am new to this. I am mainly a scientist interested in building a clone of an NVIDIA 4x GPU system, with one or two quad core CPU's. As a starting point, NVIDIA gives compatable boards; TYAN (S7025) ASUS P6T7 WS SuperComputer----(This is so far the best looking, although only really 4 GPU's, not the 7 boasted) AsRock X58 SuperComputera Now I am looking for the cheepest sales I could find on these as I want to put a system together, and would be purchasing everything seperate. Are there any suggestions, web sites, venders to look up, etc...or any other suggestions which might be usefull? From what I can tell, graphics people tend to know where things are more than most other computer persons I have met, I guess due to the competative nature of such things. My end uses would differ significantly however, as I am mainly looking for computation speed in processing rather than graphics acceleration or rendering (although the later are hand in hand). On top of this, any non-NVIDIA equivalents are appriciated, as they run the advertisments and are hard to get through (Find anything other than) on any internet searching. Grüsse, Stephan
Hello Stephan, I'll try to answer a few questions for you. However, my company is now doing commercial CUDA development and I'm under a very strict NDA, so am somewhat limited as to what I can say. We've recently started to look at OpenCL. The ATI implemtation is very young, only recently released. The nVidia OpenCL implementation (if I understand correctly) is nothing more than a wrapper around CUDA and therefore not as fast as directly using CUDA. Although, (and this is secondhand), it is more stable than the ATI implementation. I'm not sure I'd go the OpenCL route right now. Maybe next year. It reminds me very much of the Java promise of write once, run everywhere. Then you find out that there isn't a generic wrapper library that can enumerate all OpenCL devices. ie. you need to compile and link to the nVidia libs to enumerate the nVidia devices and the ATI libs to enumerate the Radeon devices. Then there is the function name mangling. You'd think that as long as the function sigs are the same you could link against one implementation and use another at runtime. Someone needs to bang heads together so the same calling convention is used by all the vendors libs. Might be worth sending an email to Gianni. The lab behind GPUGRID were hiring. And even if not, he might be able to point you in the right direction of someone who is or even talk to you about the work they are doing to port their CUDA application to OpenCL. gianni.defabritiis_at_upf.edu I'd stay away from the ASRock board. You get what you pay for. The P6T7 is a good choice for a single CPU solution. The Supermicro X8DTH is also an option for 4x double spaced 16x PCI-E slots if you are thinking about dual CPU boards. The advantage with both the Tyan and Supermicro dual CPU boards is that because they both have dual 5520 IOH's, they have enough PCI-E lanes to supply a full 16 to each of the 16x slots. Not so important for folding, but the bandwidth might be an issue depending on how much data you are shuffling between host memory and GPU memory. The P6T7 uses a pair of nf200's to "multiplex" the lanes. It does add latency when more than a single card is connected to the same "switch" . Again, how much of a disadvantage this is depends on how much data you are shifting between host memory and the cards. If you'd like to discuss further, (and it's not specifically folding related), I'd suggest you fire me an email.