Message boards :
Number crunching :
Ryzen and Threadripper
Message board moderation
Previous · 1 . . . 33 · 34 · 35 · 36 · 37 · 38 · 39 . . . 69 · Next
Author | Message |
---|---|
Ian&Steve C. Send message Joined: 28 Sep 99 Posts: 4267 Credit: 1,282,604,591 RAC: 6,640 |
One question - has anyone tried, in a controlled way, to see what happens if the GPU has too little memory to hold both tasks concurrently? I'm guessing memory conflicts and consequent data corruption, but I could be a mile out. I think it would only be an issue on 2GB cards, which is the lowest VRAM size supported on the special app anyway, cards like 750ti, 950, 1050 might not work well. I do not know what happens when you try since I do not have any 2GB cards. each Special App instance uses ~1300MB of VRAM on my host, so 2 of them ~2600MB used per GPU. another thing to keep in mind is that is uses double the system memory also. ~600MB per Special App instance. my 7-GPU system with the mutex build uses about 10GB of the 16GB installed. my 10-GPU system with the mutex build uses about 17GB of the 32GB installed. probably not an issue for most people if you have 4 or less GPUs and at least 8GB memory. Seti@Home classic workunits: 29,492 CPU time: 134,419 hours |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
Thanks Juan. AFAIK There are no one who test with less than 2 GB on the GPU . W3Perl has some 2GB 750 Ti and i believe works fine with them. Provided you not run another heavy GPU memory use application at the same time of course. He could answer the question for us. The main problem is with the host memory itself. Each WU uses about 0.5 GB of additional main memory to run so in a 4 GPU, running 6 CPU WU like my host uses about 10 GB to run including the OS itself, etc. IMHO the big problem happening on large mining hosts who normally runs with relatively small memory and has few memory slots. look this commentary from Tbar about in the thread you can't read: By my calculations 28 tasks would need around 19GBs of Memory. I suppose with only 16 Real and 6GBs SSD swap, at some point it's bound to run out. A quick check shows it would cost a minimum of around $140 for 32GB of RAM for a Two slot board. I'm not sure it would be worth it. But that not affect the most "common" 1 or 2 GPU hosts with 4 or 8 GB. One interesting point the tests showed is: slower the CPU or the HDD, bigger the gain since it will require more time to load the WU before the crunching happening. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
Answering the memory question. There is the answered received from W3Perl during the test phase: My post: So unless i'm wrong the mutex builds must work on a 2 GB, n the 1 GB i'm probably no, but need to test. His Answer: yes it runs on a GTX 750 Ti (2Gb). But sometimes there is no more 2 Gb free, Xorg may claims lots of memory. So I must set gpu_mode to never when not enough memory is found. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
Maybe a moderator could change this mgs to a more appropriated thread. |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
Maybe a moderator could change this mgs to a more appropriated thread. Opps. Sorry, I did kick off a long digression from the central topic of this thread. And I agree. It should spilt off intoa "Mutex for dedicated seti crunchers thread. Tom A proud member of the OFA (Old Farts Association). |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
Cpu crunching is never good efficiencywise for your wallet. The only cpu that comes near is the new Threadripper with 64 cores. But the price is another thing. As usual its almost always better to have a latest gen Nvidia bang/buck wise. I know we have some participants in this thread with 3900x cpus. Is anyone running a 3700x cpu.? That would be a pretty direct power draw comparison between it and a2700. I am ignoring the addition performance gains from using the AVX specific version of the app. Tom A proud member of the OFA (Old Farts Association). |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
One question - has anyone tried, in a controlled way, to see what happens if the GPU has too little memory to hold both tasks concurrently? I'm guessing memory conflicts and consequent data corruption, but I could be a mile out. The first WU loads and start to cruch as normal, the second WU stops with the task posponed error, waiting for GPU memory, then when the first one ends it restart. But in this case there are no gain in the use of the Mutex build since the second WU is not waiting on the memory ready to be crunched. On the contrary all slows down because that posponed error. It`s about the same what happening if you try to run several instances of the SoG builds on a low memory GPU. I not see any data corruption or memory conflics on my test. If the host main memory is to low, the program simply not load exactly as happening with the common builds. |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
The TDP of a 3700X is 65watts. The TDP of a 2700X is 105watts. So I expect the 3700X to be more power efficient at the same clocks. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
The TDP of a 3700X is 65watts. The TDP of a 2700X is 105watts. So I expect the 3700X to be more power efficient at the same clocks. TY. I still yearn for "more cores" but if I continue to be on the hunt to lower my power bill without lowering my crunching this looks like a possibility. Last month I paid $485 for my power bill. This month it looks like may $200. A proud member of the OFA (Old Farts Association). |
jsm Send message Joined: 1 Oct 16 Posts: 124 Credit: 51,135,572 RAC: 298 |
And crunching with the 2990WX dedicated costs £1 (GDP) per day or £365 per annum. I wonder what a 3990WX (8 way memory a must) will set me back? jsm |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
And crunching with the 2990WX dedicated costs £1 (GDP) per day or £365 per annum. I wonder what a 3990WX (8 way memory a must) will set me back? While there is no way to tell until we get "official" #'s my impression is the 3990wx runs way hotter than the 2990wx so I am guessing it will probably pull more power. Tom A proud member of the OFA (Old Farts Association). |
StFreddy Send message Joined: 4 Feb 01 Posts: 35 Credit: 14,080,356 RAC: 26 |
The TDP of a 3700X is 65watts. The TDP of a 2700X is 105watts. So I expect the 3700X to be more power efficient at the same clocks. I think these TDP values are true, when core performance boost is disabled. When core performance boost (not PBO or AutoOC!) is enabled, core package power rises significantly: my 3900x eats 145W with core performance boost enabled (without PBO and AutoOC). The TDP is 105W according to AMD. (if you do manual overclock without undervolting, power consumption will rise further) With -0.064V undervolt max core package power decreased to ~130W under World Community Grid. Noctua NH-d15s can keep this beast around 75-77 Celsius under 75% WCG load, which is fine I think. If you disable core performance boost, your CPU will operate only on its base clock and won't boost. You will loose performance, but its package power consumption will be close to the advertised TDP value. |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
Thank you for the numbers. It almost sounds like if I was running a 3900x with no turbo boost just letting it run up to its normal maximum that I might be able to pull the same power that my 2700 is now pulling and still gain 8 more threads..... More points to ponder. I am hoping to make a decision either during the Black Friday/Cyber Monday sales cycle or probably for Christmas. Somehow I don't really expect anyone to offer a 3900x for less than list price! But who knows, maybe the 3950x will have hit the shelves and started driving down the price of the 3900x ;) Tom A proud member of the OFA (Old Farts Association). |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
If you disable core performance boost, your CPU will operate only on its base clock and won't boost. You will loose performance, but its package power consumption will be close to the advertised TDP value. You can also just set a manual all-core overclock you are comfortable with. And set the necessary voltage to run that. The plus side of doing that is once you set a manual multiplier the Vcore starts out at a default low value of 1.017V. Which runs very cool. You just have to add in some positive offset to get your set clock to run stable at the minimum required voltage. Which generates less heat and uses less power. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
Another great article from 1usmus, the developer of the Ryzen Memory Calculator software. He has prepared another article at TechPowerUp about his new 1usmus Custom Power Plan for Ryzen 3000 Zen 2 Processors. https://www.techpowerup.com/review/1usmus-custom-power-plan-for-ryzen-3000-zen-2-processors/ I suggest anyone running a Zen2 3000 or better processor as well as anyone considering the new upcoming Threadripper processors to give it a read. Specially so if you run Windows10. He discusses the shortcomings of the Windows thread scheduler as well as the lame attempt by AMD to boost cpu clocks to advertised boost speeds with the recent AGESA 1.0.0.3 ABBA BIOS update. I would like to hear from anyone who installs his AMD Power Profile in Windows10 whether they see any improvements or the advertised improvement that 1usmus espouses. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Jord Send message Joined: 9 Jun 99 Posts: 15184 Credit: 4,362,181 RAC: 3 |
It was a bit of a wait, but I finally snatched an AMD Ryzen 9 3900X CPU tonight, just moments ago. Can go fetch it tomorrow. Will now look for a motherboard. Pref at the same store of course. :) (Edit: I think I'll go for an Asrock X470 Taichi, it may just need the BIOS update, but the store can do that for me. Have to order the board anyway) I do think I paid a higher price for it than when it was first released, but still under 600 euros (which is main stream price, I see): €592,95 |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
Congratz, Jord. Think you will like it. Plenty of cpu threads to play with. The math performance upgrade of Zen 2 is really obvious over Zen+. Happy you could find the cpu at close to original MSRP. Though I still see the cpu listed at Newegg.com for above MSRP but not at the absurd prices that were listed by third party sellers when the cpu was not in good supply. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Jord Send message Joined: 9 Jun 99 Posts: 15184 Credit: 4,362,181 RAC: 3 |
I'd just planned on going for the €339,- 3700X when I saw that one of the stores I frequent had one 3900X... didn't wait, didn't think it over, didn't sleep a night over it, just plain went for it. :) So now that that's out of the way, next is the videocard. Most all RX5700 XT's are plain ugly. And Nividia with their non-Super and Super idiocy isn't making things easier. Go for a 1650, 1650 Super, 1660, 1660 Super, 2070, 2070 Super or plain 2080? Go away Nvidia! Make things easy! ;-) |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
I think you have to stay away from Navi cards still with the driver issues with OpenCL not computing results correctly. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Wiggo Send message Joined: 24 Jan 00 Posts: 36779 Credit: 261,360,520 RAC: 489 |
And Nividia with their non-Super and Super idiocy isn't making things easier.That's really easy, the Super models are performance enhanced replacements for the non-Super models. ;-) Cheers. |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.