mersenneforum.org Radeon VII @ newegg for 500 dollars US on 11-27
 Register FAQ Search Today's Posts Mark Forums Read

 2020-05-31, 17:59 #276 Uncwilly 6809 > 6502     """"""""""""""""""" Aug 2003 101×103 Posts 23·52·43 Posts That is one ugly beast.
 2020-05-31, 19:05 #277 dcheuk     Jan 2019 Pittsburgh, PA 3×7×11 Posts I have 2 Radeon VIIs that was purchased at the same time from the same vendor. I assume the cards came off different assembly lines since one of the memory is made by Samsung while the other by Hynix. The Hynix card is able to hit 1600Mhz, 900mV, 1200MHz using only 150W with junction temperature consistently below 90C (indoor temperature 65F), while each PRP test rarely any errors. Meanwhile the Samsung card throws black screen whenever I set the memory clock above 1050MHz or gpu clock above 1500MHz. It almost always runs 5-10 degrees hotter and uses 5-10W more, under the same settings and always return 3-10 (sometimes over 20) errors. Tried swapping their slots, rolling back to 19.4.2 or earlier drivers, using only one GPU at a time, experimented on both 9700k or 3700x. Nothing really changes. We probably just have to accept that some cards are just badly made compared to others I guess. Maybe this samsung card is just a lemon, but I was surprised the Hynix card performed significantly better. I have sold the samsung card to a friend that only play games at a discounted price and he couldn't care less lol.
2020-05-31, 19:18   #278
ewmayer
2ω=0

Sep 2002
República de California

2×13×443 Posts

Quote:
 Originally Posted by Uncwilly That is one ugly beast.
That depends on what one finds beautiful in such a thing. :) The typical setup inside an ATX tower case system with multi-GPUs ain't any prettier.

Update on the power usage - I had been running both cards 1 and 3 (the 2 rightmost in the pic, my numbering is based on the order they appear in the ROCm listing) at sclk = 3, but card 1 was able to handle mclk = 1200 whereas card 3 was not, had to dial the latter down to 1150. I noticed, though, that the timings differed by maybe at most 1%. Dialed card 1 down to 1150 as well, that cut 20W off the usage, with the stated 1%-ish timing hit. So did same to card 2 (which had been running at sclk = 4, mclk = 1200), another 20W cut, and that card - which was marginal @mclk=1200, at that point it was ~20% through a PRP run @5.5M FFT, with 18 G-check errors for job1 and 19 for job2 - also runs much more stably at the lower mem-clocking ... over 12 hours after the mclk fiddle, no addition errors for the 2 runs, both now ~60% done.

TDP now ~750W, still a bit on the high side for my PSU, but better.

So the general idea is: Use the sclk setting as a coarse dial on performance and wattage, and mclk as a fine control.

Last fiddled with by ewmayer on 2020-05-31 at 19:19

2020-05-31, 19:42   #279
moebius

Jul 2009
Germany

22×5×17 Posts

Quote:
 Originally Posted by Uncwilly That is one ugly beast.

MOBILE OUTDOOR HEATING POWER PLANT...

2020-05-31, 20:36   #280
Prime95
P90 years forever!

Aug 2002
Yeehaw, FL

23·311 Posts

Quote:
 Originally Posted by ewmayer Update on the power usage - I had been ... mclk = 1200 .... Dialed card 1 down to 1150 as well, that cut 20W off the usage, with the stated 1%-ish timing hit.

I'm try to replicate with less success. Rocm-smi shows 3% power reduction (from ~170 to ~165) with an unknown speed penalty (still measuring).

Perhaps the 20W savings you see are 5W from the Radeon plus 15W from the PSU being more efficient.

BTW, for anyone else thinking of replicating your setup a platinum power supply will likely pay for itself in a reasonable time frame.

Edit: Runs 1.7% slower

Last fiddled with by Prime95 on 2020-05-31 at 23:05 Reason: added timing

 2020-05-31, 20:56 #281 pinhodecarlos     "Carlos Pinho" Oct 2011 Milton Keynes, UK 111468 Posts Knowing the accuracy of the meter is also important to understand your avoided 20W claim since it might not be representative.
2020-05-31, 22:29   #282
ewmayer
2ω=0

Sep 2002
República de California

2·13·443 Posts

Quote:
 Originally Posted by Prime95 Was this savings measured at the wall or using rocm-smi? I'm try to replicate with less success. Rocm-smi shows 3% power reduction (from ~170 to ~165) with an unknown speed penalty (still measuring). Perhaps the 20W savings you see are 5W from the Radeon plus 15W from the PSU being more efficient. BTW, for anyone else thinking of replicating your setup a platinum power supply will likely pay for itself in a reasonable time frame.
At wall, using the brand of meter I linked in my final Bill of Materials. Also needs to be roughly time-averaged ... before the 1200->1150 mem-downclocking of cards 1 & 2, the meter had been showing wattages in a narrow (+- 5W most of the time, rarer +- 10W swings) band right around 800. After, similar narrow band centered at 760W.

I'm planning to lay in a backup PSU over next couple months, so any good deals in the 850-1100W/gold-platinum range people find online, please post here or PM me.

Coming weeks mini-projects are final stuff - homemade 3x80mm fan guard (the R7 fan holes are actually a smidge wider than that) for end-of-board card 2, and silicone-adhesive to glue the feet of the glass-topped monitor stands I bought. One thing I only realized after the latter arrived - I got both with smoke-color glass, black-painted on the underside ... should've got the top one sans the black paint, to allow the Lights of Vegas, erm I mean the R7 LEDs, to show through. I scraping a little at one of he corners of the black paint to see if it can be removed that way - you figure glass being so smooth any paint should come off pretty easily, but whatever they used is very tough, almost like black the anodization one sees on metal. Gonna pick up a small can of paint stripper tomorrow, again try a small dab in a corner, to see if that works. If not, ah well. (I know, vanity - sounding like a gamer who wants all his buds to ooh and aah over his new rig here. :)

2020-06-01, 04:23   #283
LaurV
Romulan Interpreter

Jun 2011
Thailand

873610 Posts

Quote:
 Originally Posted by ewmayer Meet The Gimp ...
That is a nice build, however I have a question about that photo: where da fun is the second PS? (850 looks a bit too less to keep those 3 beasts running max speed)

edit:
Quote:
 Update on the power usage
sorry, didn't see that (page turned, your pic was the last on the former page). Anyhow, you need a larger PS, for sure! You may even save power (due to a better efficiency of the PS not working close to its limit). But that you must check first...

Last fiddled with by LaurV on 2020-06-01 at 04:27

 2020-06-01, 13:34 #284 sdbardwick     Aug 2002 North San Diego County 23·5·17 Posts I wouldn't replace the PSU unless more efficiency is desired. The Corsair RM850 is a quality PSU with honest ratings; 850W @ 50C ambient. If the draw is 760W at the wall, the rig is pulling about 675W DC, or about 79% of rated capacity (assuming 88.5% efficiency). A more efficient unit might make fiscal sense, but one must investigate the actual efficiency curves for replacement candidates; the old "max efficiency at 50% of rated power" rule cannot be relied upon. Many supplies peak at lower loads, but the slope after the peak can be gentle.
 2020-06-05, 21:40 #285 ewmayer ∂2ω=0     Sep 2002 República de California 2·13·443 Posts Couple days ago finally moved the beast into its permanent home, a pair of 4"-high glass-top monitor stands, bottom one flipped upside down and foot-bottoms epoxied together to make an 8"-high air gap, test-frame holding best sits on bottom glass, monitor on top, looks great, but air now constrained to escape out sides and front, upward and back are blocked. On top of that we were in midst of a 4-day heat wave, so that required me to up all the GPU fan settings, downclock the middle GPU further to sclk = 2 (which incurs a major 15% perfromance hit), and also blowing a tower fan (same one I use to keep cool @my desk) at the unit in this heat. In the meantime ordered a compact high-CFM Honeywell desktop fan, a corded one, not one of those wimpy USB fans which cost just as much and blow less than 1/10th as much air. That arrived yesterday and now sits just at the left end of the Glass House and blow air sideways across the hardware. To make room for it on my desk, I moved the cellphone charging station that had been in that spot under the left end of the glass enclosure as shown, the square USB fan that had been colling that is no longer needed, and the much-more-powerful airflow from the new fan cut all my Mlucas runtimes on those phones by ~10%. The fan makes a massive difference in terms of GPU cooling - was able to cut the GPU fan settings across the board, and the net effect is much quieter. Hot weather gone at last as of yesterday evening, revealing one more plus from the new arrangement - next time we have a heat wave, I won't need to fiddle the individual GPU fan settings upward, I'll just click the knob on the desk fan from its current '1' (minimum) setting up to '2', where I had it during the last of the heat yesterday. And best of all, I finally have some semblance of my desktop back! Attached Thumbnails   Last fiddled with by ewmayer on 2020-06-05 at 21:40
 2020-07-10, 19:03 #286 ewmayer ∂2ω=0     Sep 2002 República de California 1151810 Posts Yesterday scored a pair of like-new R7s from an ebay seller, $1450 incl ship, but CA sales-gouge adds 8.5%, or$123. One will go into my Haswell system thus completing my build-out (no more room or PSU watts-to-spare in either of my 2 rigs), the other I plan to offer to fellow GIMPSers at cost plus ~$15 prio-mail shipping (within US), or$800. PM me if interested. More than I'd hoped to pay, but I guess my first 4 R7s spoiled me as to price. :) Still need to take delivery and test 'em both out, obviously.

 Similar Threads Thread Thread Starter Forum Replies Last Post ET_ GPU Computing 1 2019-07-04 11:02 M344587487 GPU Computing 10 2019-06-18 14:00 jasong GPU Computing 0 2016-11-09 04:32 0PolarBearsHere GPU Computing 0 2016-03-15 01:32 firejuggler GPU Computing 33 2014-09-03 21:42

All times are UTC. The time now is 07:33.

Wed Sep 23 07:33:04 UTC 2020 up 13 days, 4:44, 0 users, load averages: 2.80, 2.52, 2.31