Questions and Answers :
Unix/Linux :
Maybe off topic but......
Message board moderation
Previous · 1 · 2 · 3 · 4 · 5
Author | Message |
---|---|
Send message Joined: 12 Nov 21 Posts: 236 Credit: 575,038,236 RAC: 0 |
Well ATI/AMD gpus have always needed WAY less cpu support for gpu tasks compared to Nvidia cards. Generally, the accepted consensus opinion is that Nvidia cards need a full cpu core/thread to support gpu tasks. Yep. My NVIDIA P1000(not exactly a screamer) is running 30% CPU, at close to 100% GPU utilization. The Sapphire R9 280X is running 3 concurrent tasks. CPUs are 6.7%, 6.8%, and 9.4. This occasionally bumps up to 30% when a task starts up, but typically only bumps 20% or so. |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
Well ATI/AMD gpus have always needed WAY less cpu support for gpu tasks compared to Nvidia cards. Generally, the accepted consensus opinion is that Nvidia cards need a full cpu core/thread to support gpu tasks. Well, I have finally maxed out what I think is reasonable for my systems. On the AMD/ATI systems, I am running .15 CPU and .20 on GPU. This gives me all of my CPUs running and approx. 20 minutes per WU on 5 GPU WUs. On the Nvidia systems, I am running .15 on the CPUs and .33 on the GPUs. This gives me all of my CPUs running and approx. 20 minutes per WU on 3 GPU WUs. I was really amazed that I could set the CPUs so low, but it's been a day and a night with no problems. All of this running on air and my fans are about 30%. Naturally, my load on all is 100%, although GPUs sometimes drop a little below that. |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
Keith, Another question: I have been wondering if I can put two of my RX 560's in one machine. I have a RX 570 that I could put in the other machine. Also, what about putting two 1050ti's in one machine? Non of them require the connecting wires between the cards. Just thinking...... Thanks, Allen |
Send message Joined: 13 Oct 21 Posts: 44 Credit: 226,275,245 RAC: 11,113 |
I think AMD GPUs are better suited for MilkyWay because of their high native double precision capability compared to NVIDIA cards. This is probably also the reason that their CPU usage is much less. I recently got a R9 280X (a 9 year old card) and was amazed at its capabilities in this project compared to the 3060Ti (latest gen. card) that came with my PC. After some testing I found that for most productivity I can run 4 concurrent GPU tasks (.25 GPU, .24 CPU settings). The average time per tasks is about the same as one task on 3060Ti running one at a time. Running more than 1 at a time on 3060Ti is counterproductive as the average time per task slows down too much. It'd be interesting to try NVIDIA TITAN V here as its rated double precision capability is about 7 times that of an R9 280X. Even at 5 years old, its price is about 16 times that of a R9 280X though. I think a good modern consumer grade card that'd run well on this as well as other projects is AMD RX 6800 XT. Much cheaper than 6900 XT and only ~10% less capable, with excellent single precision and very good double precision ratings. |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 5 |
Keith, Yes you can add as many gpu's to your machine as it can handle, one guy has 7 in one machine!! You CAN mix and match the gpu brands but if you don't have to it might be easier with the drivers. The connection cable is NOT helpful in crunching, in fact it can slow things down so unless you game or do heavy duty image stuff that takes advantage of your multiple gpu's leave it in the drawer. |
Send message Joined: 13 Apr 17 Posts: 256 Credit: 604,411,638 RAC: 0 |
As already said elsewhere, for running more than one GPU you only need to put "<use_all_gpus>1 ..." into cc_config file. |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
Thanks to all on the GPU cards issue. I thought you must be able to, but wasn't clear on it. Now the question is, do I need to be careful not to overload my Power Supply? I guess I do. I have 750 Gold in one of them, but not sure about the others. More than likely, only around 500 Watts on the rest. Thanks! |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 5 |
Thanks to all on the GPU cards issue. I thought you must be able to, but wasn't clear on it. Yes you need to be very careful about that or you could pop it!! |
Send message Joined: 24 Jan 11 Posts: 712 Credit: 552,280,121 RAC: 47,089 |
Keith, Sure, why not? As long as the cards can fit on the motherboard, preferably directly in the PCIe slots, go for it. And you have enough PCIE power cables to feed them and sufficient power capacity from the power supply. If you two or more of the same type card in the host, BOINC will crunch on both of them as long as they are detected at startup. If you put dissimilar cards into the host, then BOINC will only crunch on the most capable card. In your example, then only the 570 will get detected and crunched with and the 560 will be ignored. To get around that problem you need to set a cc_config.xml parameter in the <Options> section. <use_all_gpus>1</use_all_gpus> and re-read config files in the Manager or restart the client and Manager to get BOINC to crunch on both the 570 and 560. |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
I think AMD GPUs are better suited for MilkyWay because of their high native double precision capability compared to NVIDIA cards. This is probably also the reason that their CPU usage is much less. I recently got a R9 280X (a 9 year old card) and was amazed at its capabilities in this project compared to the 3060Ti (latest gen. card) that came with my PC. After some testing I found that for most productivity I can run 4 concurrent GPU tasks (.25 GPU, .24 CPU settings). The average time per tasks is about the same as one task on 3060Ti running one at a time. Running more than 1 at a time on 3060Ti is counterproductive as the average time per task slows down too much. It'd be interesting to try NVIDIA TITAN V here as its rated double precision capability is about 7 times that of an R9 280X. Even at 5 years old, its price is about 16 times that of a R9 280X though. I think a good modern consumer grade card that'd run well on this as well as other projects is AMD RX 6800 XT. Much cheaper than 6900 XT and only ~10% less capable, with excellent single precision and very good double precision ratings. I certainly agree with you on both GPU's..I can only tell you that for me, the advantage of running more units per GPU on the Radeon just kept getting better until at least 5 per GPU. After that it didn't seem to make much difference. |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
Keith, Thanks again. I don't have to worrying about the 6 pin cables on the 560's since they don't use them, so I should have enough slots to take care of everything. |
Send message Joined: 24 Jan 11 Posts: 712 Credit: 552,280,121 RAC: 47,089 |
I don't know what times a single task on a AMD can run because I am sure most AMD users run at least 2x, 3X, 4X or more. I can run a single Separation task in under a minute. I run 2X on all my cards. Certainly better for the 3000 series compared to previous generations. So blanket statement that Nvidia cards are no good at Milkyway does not hold water. |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
I don't know what times a single task on a AMD can run because I am sure most AMD users run at least 2x, 3X, 4X or more. I can run a single Separation task in under a minute. I run 2X on all my cards. Certainly better for the 3000 series compared to previous generations. So blanket statement that Nvidia cards are no good at Milkyway does not hold water. Did someone say that the Nvidias are not good for MW? I think they run fine, I just can't run as many units at one time (with an advantage or time) than I can on my 560's. I run 3 on the Nvidias and 5 on the 560's with about the same 20 minutes per task. One task on the 560 is about 6+ minutes and 7+ minutes on the Nvida. One thing for sure, they sure do generate a lot of heat!! |
Send message Joined: 24 Jan 11 Posts: 712 Credit: 552,280,121 RAC: 47,089 |
There is a recurring theme posted here at Milkyway and throughout the BOINC forums that you don't want to use a Nvidia card at MW and you should always purchase a AMD because of their superior FP64 calculation rate. Best AMD card at MW has been the Radeon VII. The 3080 Ti matches and beats that card. No arguments about the heat production from Nvidia cards though as I look at my power bill every month from the A/C usage to try and remediate the heat output from the computers with 13 cards. |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
There is a recurring theme posted here at Milkyway and throughout the BOINC forums that you don't want to use a Nvidia card at MW and you should always purchase a AMD because of their superior FP64 calculation rate. Ya, it's tough in the Summer bills, but I don't mind it at all in the Winter. I've got low voltage GPU's. I can't even imagine how much heat and cost you have with the GPUs you're running. Whew! |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
There is a recurring theme posted here at Milkyway and throughout the BOINC forums that you don't want to use a Nvidia card at MW and you should always purchase a AMD because of their superior FP64 calculation rate. Hi Keith, Just thought I would drop you a line and let you know that I finally got my 580 running on the Linux machine. It's running 10 wu's at about 14 minutes per. The 560 was running 5 wu's but was taking about 20 minutes per. Now I've taken the other 560 and doubled up on one of the Windows machines and get 10 wu's at a time in about 20 minutes. Seems that the 580 is a monster compared to the 560, who knew?? How are things with you? Any new upgrades? Allen |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 5 |
There is a recurring theme posted here at Milkyway and throughout the BOINC forums that you don't want to use a Nvidia card at MW and you should always purchase a AMD because of their superior FP64 calculation rate. The difference IS huge as compared to the compute cores for each one AMD 560 896 AMD 580 2304 |
Send message Joined: 2 Mar 20 Posts: 131 Credit: 319,359,639 RAC: 17,052 |
Right you are! I never even bothered to consider the difference between the compute cores. I need to start paying attention. I've got a 570 sitting on the shelf. Time to get it up and running too. Thanks mikey!! |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 5 |
Sure no problem.the 570 is also a 1/16 FP64/FP32 ratio and 2048 compute cores so it should do very well as well. |
©2024 Astroinformatics Group