Posts by QuintLeo

41) Message boards : Number crunching : amd 480 and opencl issues (Message 7677) Posted 20 Jun 2017 by QuintLeo Post: This project should work with an Intel iGPU - the basic DNet client does so, but the performance is really really bad - almost bad enough to not bother with it. 150 MKeys/sec on 630 graphics (G4600) if you're not doing ANYTHING else on the system, commonly more like 90 if you ARE using the CPU for anything - my generations-old A10-5700 manages a bit over 400 WITH the CPU doing stuff. For reference, you have to go to the higher-end HIGH COST workstation CPUs from Intel to get a "better" iGPU than the 630 (it's the same iGPU that's on the Kaby Lake i7-7700k among other Intel offerings).
42) Message boards : Number crunching : best settings for crunching (Message 7676) Posted 20 Jun 2017 by QuintLeo Post: For any GPU, maximum core clockrate you can run stably, minimum memory speed you can set your card to, and you can probably get away with lowering the TDP limit quite a bit (depends on how far you can go and still get maximal boost clock).
43) Message boards : News : Disk failure on the main project server (Message 7639) Posted 3 May 2017 by QuintLeo Post: IMO server should be running on a RAID.
44) Questions and Answers : Unix/Linux : opencl problem in dual gpu(amd and nvidia mix) (Message 7621) Posted 17 Apr 2017 by QuintLeo Post: BTW - AMD APU and AMD discrete card are trivial, if you have a motherboard with an updated-enough BIOS to recognise the AMD discrete card OR if you have a recent-enough APU. I've had SEVERAL such systems running at times, though I do recall having issues with my A10-5700s not being usable when I installed a discrete card - 'till I upgraded the BIOS on the motherboards in question. Sadly, those motherboards STILL won't let me use the GPU on my A10-5700s and my RX 470s at the same time - but they work fine with my older HD 7xxx and R9 2xx cards now.
45) Questions and Answers : Unix/Linux : opencl problem in dual gpu(amd and nvidia mix) (Message 7620) Posted 17 Apr 2017 by QuintLeo Post: It's entirely possible to run both an AMD and a NVidia card on the same machine under LINUX - though it's a non-trivial setup (one of the FEW things Windows does better than LINUX). The main limit though is that you can only get display output from ONE brand of card on the same machine, AFAIK, without doing some sort of "switcher" application that reconfigures things then reboots the machine. You CAN get both to crunch at the same time - best to use an AMD card as your display output card though, as there is an option on the Nvidia driver installer to install it so it does NOT install OpenGL and end up overwriting some of the stuff the AMD driver installs - the AMD driver doesn't have that option. Unfortunately, I can't find my note with the command-line option to use (I moved recently and am still getting stuff unpacked and sorted out) - I DO recall you have to manually install the Nvidia driver AFTER you have installed the AMD driver.
46) Questions and Answers : Windows : Attempting anonymous platform for Intel GPU (Message 7619) Posted 17 Apr 2017 by QuintLeo Post: For reference, my A10-5700 (384 shaders, runs the GPU side stock at around 750 Mhz TeraScale 4 IIRC)) does right about 400 Mkey/s on core 3 - which is a much more fair comparison to an Intel GPU on a CPU setup. A10-7860K or A10-7890K (512 shaders, running the GPU overclocked a bit at 800 Mhz, GCN) both achieve a hair over 600 Mkeys/sec (which happens to be the same figure I see out of my HD 7750 cards that are also 512 shaders at their stock 800Mhz clock). I have been curious about what kind of keyrate the Intel graphics-on-the-CPU stuff was capable of. I'm afraid I have to say that I am not impressed, though given the other competative benchmarks I've seen I'm not exactly shocked. Intel should just give up on trying to compete with AMD or Nvidia on graphics, they've never been close and usually aren't even in the same ballpark. BTW - RC5 work doesn't involve floating-point at all, so trying to compare keyrate to FLOPS makes ZERO sense anyway.
47) Questions and Answers : Windows : R9 270x failing (Message 7618) Posted 17 Apr 2017 by QuintLeo Post: I have run MooWrapper on much older and lower-end GPUs than your R9 270x (including the GPU part of a pair of A10-5700s and a bunch of HD 7750s), as well as on as high as the R9 290. Try driver version 15.12 - that seems to be the best performance highest stability driver for any AMD GCN cards prior to the Fury. I suspect however that you have a configuration issue somewhere.
48) Message boards : Number crunching : AMD RX 480 (Message 7617) Posted 17 Apr 2017 by QuintLeo Post: Distributed.net RC5-72 speeds have almost ZERO resemblence to any other benchmark in existance, don't bother trying to compare. The only benchmark I have ever seen that is even CLOSE to similar is Bitcoin mining, which is also a form of cryptographic work in it's root basis - but still involves a lot more data usage and therefore memory throughput affects it a lot more than RC5-72 work does. The keyrate of any AMD GPU that is capable of running Dnet at all is almost 100% proportional to the product of the number of shaders/cores on the card time the core clockrate. The RC5 CODE is very simplistic - it does a few rotates and a few adds and not much else - so comparing it to a benchmark that does a LOT more complicated code and needs a lot of data to/from memory doesn't work. I don't have an exact number, but it works out to 1 Megakey/sec per shader/core at a little under 700 Mhz clock - and it doesn't matter if it's a current GCN card or the older Terrascale stuff like the HD 76xx series and older (including the applicable GPU on AMD A-series APUs). The memory usage is TINY - all the "core" code AND all the data fits easily in cache memory, even on the older CPU-based clients, since at least the AMD K6 and Intel Pentium II/Celeron series which had cache memory sizes in the several KILObyte range. It is entirely practical to clock your memory on a GPU to the LOWEST it will go and have ZERO effect on keyrate, while saving a bit of power and heat. RX 480 has 2304 cores, the R9 290X has 2816 (22% more) - the RX 480 DOES clock higher but it is AT BEST going to have similar keyrate - it comes down to can you clock the RX 480 more than 22% higher than you managed on your R9 290X. On the POSITIVE side, the RX 480 uses a LOT less power to manage similar keyrate. I don't personally have a R9 290X or a RX 480, but I DO have several R9 290 and a pair of RX 470 - which are both slightly-downsize from your cards, and have the same "about the same keyrate due to more cores but slower clock" comparison you see.
49) Message boards : Number crunching : Multiple GPUs on same WU (Message 7616) Posted 17 Apr 2017 by QuintLeo Post: As far as I know, MooWrapper specifically runs each work unit on an individual GPU, even when you have multiple cards of the same model in the machine, and there is no viable way around this (technically there IS a potential work-around via editing the "master" copy of the .ini file but it messes up the wrapper assign work code if you try to do so). If you're using a mixed-card machine, this actually makes a LOT of sense as it prevents one WU tying up all your GPUs while there is only enough left for one to work on, while having little if any downside on a machine with all of the GPUs in it the same. The base DNet client itself does not assign one key to multiple GPUs to work on at the same time, even when running it as a multi-GPU setup it still assigns a key to a single GPU, so there is no actual benefit to trying to force MooWrapper to assign one WU to all the GPUs in your system.
50) Message boards : Number crunching : AMD RX 480 (Message 7501) Posted 22 Nov 2016 by QuintLeo Post: The RX480 is inferior in memory bandwidth (though that shouldn't be an issue) to the R9 290X. It has fewer cores, but clocks them higher. Also, the current 520 client being used for opencl support doesn't work all that well with cards like the RX 480 that use opencl 2.0, especially under LINUX - it really should get upgraded to the 521 client that DOES.
51) Message boards : Number crunching : WU take the wrong GPU ! (Message 7500) Posted 22 Nov 2016 by QuintLeo Post: Is this under LINUX? Moowrap has a long-running issue where it is flat out IGNORING the "ignore_nvidia_device" configuration option in cc_config.xml under LINUX, but it works under Windows. It is also ignoring what you tell it on the webpage configuration for LINUX installs. The version of the 520 client it is using for opencl under LINUX does NOT appear to support openCL 2.0 - that client needs to get updated to the CURRENT AMD64 OpenCL 521 client ASAP.
52) Message boards : Number crunching : r9-280x and r9-290 issues (Message 7499) Posted 22 Nov 2016 by QuintLeo Post: 15.12 drivers work well with the R9 290(x). The 16.x series seem to not work as well, as fast, or as reliably - they're intended for the RX 4xx series. and AMD doesn't seem to bother testing them against older cards any more.
53) Questions and Answers : Unix/Linux : opencl problem in dual gpu(amd and nvidia mix) (Message 7498) Posted 22 Nov 2016 by QuintLeo Post: You USED to be able to replace the client with the 521 dnet client that supports running on a specific GPU, and set up the default .ini file to use the CORRECT GPU. This no longer works since the change to the "amd_101" wrapper - the wrapper apparently insists on REPLACING the client with the one it wants AND WILL OVERWRITE IT WITH A BROKEN 520 CLIENT THAT DOES NOT SUPPORT THE -devicenum COMMAND LINE PARAMETER. It also appears to generate the .ini file itself now instead of using one out of the moowrap.net directory. I have 3 systems with AMD APUs that can not be used on BOINC for this reason - the NVidia cards in the system are for use on Folding@Home, and if I try to run BOINC on the system it FORCES ONE OF THE NVIDIA CARDS TO BE USED INSTEAD OF THE A10 APU - and completely BLOWS OFF anything in cc_config.xml or the website checkbox stuff to do so. This makes it IMPOSSIBLE for me to run BOINC/Moowrap on these systems which is starting to tick me off a LOT, as I've wasted HOURS trying to get them working and they refuse to behave or pay attention to the BOINC configuration options that are supposed to CONTROL this sort of thing. This setup WORKS under Windows just fine, it's not acceptable for it to act completely differently AND BE BROKEN under LINUX.

Previous 20