mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware > GPU Computing

Reply
 
Thread Tools
Old 2020-08-11, 00:15   #331
PhilF
 
PhilF's Avatar
 
"6800 descendent"
Feb 2005
Colorado

52·29 Posts
Default

You could use something like this to measure the current through the cable (using the clamp-on method). From that you can compute the wattage easily enough if you want (P = I x E):

https://www.amazon.com/Multimeter-Au...7104635&sr=8-7
PhilF is offline   Reply With Quote
Old 2020-08-11, 02:41   #332
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

13·227 Posts
Default

Quote:
Originally Posted by ewmayer View Post
All feel cool to the touch - or better, not appreciably warmer than ambient. However, I don't know how much wattage the Frankencable-supplied 8-pin PIe power plog is carrying, compared to its conventionally-cabled neighbor. I figured using just a single Frankencabled plug would be the safest way to start, since if said cable is having any excess-resistance issues, it would [a] be spottable via a warm cable and [b] naturally lead to the other supply cable taking up more of the load. I suspect both-plugs-via-Frankencable will be a stiffer test in a week, though if things get warm in that scenario I have the option of splitting power to both GPUs 3 and 4 the same way as GPU 3 at present, one plug conventional 8-pin PCIe power, the other Frankencable. Are there reasonably cheap mini-watt-meters one can insert in between an 8-pin PCIe power plug and its input receptacle?
The 8 pin cables are designed to carry 150 watts while the 6 pin only 75, and the graphics cards should limit themselves to those values. You may wish to use the proper cables for the 8 pin connectors and the frankencables for the 6 pin, if possible.
Mark Rose is offline   Reply With Quote
Old 2020-08-11, 05:09   #333
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

11,743 Posts
Default

Quote:
Originally Posted by Mark Rose View Post
The 8 pin cables are designed to carry 150 watts while the 6 pin only 75, and the graphics cards should limit themselves to those values. You may wish to use the proper cables for the 8 pin connectors and the frankencables for the 6 pin, if possible.
AFAICT I am using all connectors properly - the R7 shipped with a pair of 2x-6-pin-to-1x-8-pin pcie adapters, my current Frankencable has each 2x-15-pin-sata-to-6-pin-pcie adapter feeding a 6-pin pcie power input on one of those, i.e. we should be able to drive the endpoint 150w 8-pin pcie plug at its rated power, if each of the 4 upstream sata connections can handle ~40w. As my cards are underclocked, though, I don't expect either of the two 8-pin external power inputs to the GPU to ever need to carry more than, say 100w, conservatively speaking.
ewmayer is offline   Reply With Quote
Old 2020-08-11, 05:12   #334
chalsall
If I May
 
chalsall's Avatar
 
"Chris Halsall"
Sep 2002
Barbados

101001101001112 Posts
Default

Quote:
Originally Posted by ewmayer View Post
As my cards are underclocked, though, I don't expect either of the two 8-pin external power inputs to the GPU to ever need to carry more than, say 100w, conservatively speaking.
So, you're guessing. You're inferring. You don't actually know.

Not a problem. Actually, a very common situation.
chalsall is offline   Reply With Quote
Old 2020-08-11, 06:42   #335
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

11,743 Posts
Default

Quote:
Originally Posted by chalsall View Post
So, you're guessing. You're inferring. You don't actually know.
Based on watts-at-wall, each card draws ~200w ... but I'm guessing that
the wattmeter is working properly and giving reasonably accurate numbers.

Each card has 3 power inputs - the pcie bus rated at 75w, and two 8-pin pcie plugs rated at 150w each ... but I'm guessing that the pcie bus is carrying some fraction of the power, not just data,.

Thus two 8-pin pcie plugs need supply < 200w together ... but I'm guessing that the resulting load is reasonably balanced across the 2 plugs.

And yes, as long as the guesswork is based in reasonable data and conjectures, I'm OK with that. Some people, and not just ones named Heisenberg, would say life is an exercise in serial guesswork.
ewmayer is offline   Reply With Quote
Old 2020-08-11, 22:19   #336
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

11,743 Posts
Default

[Nomenclatural note: henceforth will be using the initialism TDP to mean watss-at-wall-plug.]

Followup note re. the rocm-displayed vs wall wattages for the R7 - with my system with 3 R7s all running downclocked to sclk=3, rocm shows each consuming ~170w for a total ~500w, whereas the wall-plug wattmeter implies something closer to 650w. In fact both sets of numbers can be correct - rocm shows wattage consumed at the GPU, but we expect some losses in power generation (i.e. at the PSU) and transmission (all the cabling/connectors). This however raises the question: when PSU manufacturers tout the wattage ratings of their gear, are they talking about watts of PSU *output*, or TDP? If the former, then my drawing, say, 900w at wall could still leave me well under the rated 850w of my PSU, if the power generation step is costing > 50w.

Anyhow, system ran stably 24 hours (TDP more or less constant at ~750w) with one GPU drawing power from a mix of pcie bus, 8-pin pcie power input and power drawn from one 6-pin PSU SATA output and run through the Frankencable, so a couple hours ago shut it down, hooked up GPU 4 to conventional dual-8-pin-output pcie power cable, booted no problem, been running GPUs 1-3 as before and GPU 4 idling no problem, TDP ~770w, with the extra 20w reflecting the idling GPU 4, which rocm also shows drawing 20w.

An hour later, on to the next step - again, would prefer to be using the intended pair of Frankencables for this, but let's see if using just 1 allows us to run GPU 4 at all: set that card's sclk=1, the lowest possible, fired up a single gpuOwl job ... 1 hour later, no problem, rocm shows cards 1-4 (0-3 in rocm's zero-offset indexing scheme) drawing 170,175,165,91w, TDP has jumps to ~885w, i.e. rocm shows card 4 consuming 70w more than at idle, TDP shows total wattage up 115w due to firing up code on that card. OK, fire up 2nd job on card 4 - we don't expect the watts to increase much as result of that, we do it for the total-throughput boost resulting from running 2 jobs per card - rocm shows 97w (+6), TDP up by a similar amount, to just over 890w.

Run timings - cards 1 and 2 are running expos near the upper limit of the 5.5M FFT range, so to get as close to an apples-to-apples comparison, we compare timings on cards 3 and 4, the first of which is doing 1 expo @5.5M and 1 in the low part of the 6M range, and the second is doing 2 expos @6M. Note I haven't yet mem-upclocked card 4 like the others, since that draws more watts:

2 jobs, 5.5+6M FFT, sclk = 3, mclk = 1150: 1460 and 1480 us/iter

2 jobs, 6+6M FFT, sclk = 1, mclk = 1000: both @ 2335 us/iter

Next up the mem-clock on card 4 to 1150, same as the others (except for card 1, which has shown it can handle mclk = 1200 stably) ... we need to be careful to immediately reset sclk to 1 on card 4 after this, because the mclk-fiddle resets those to default (as in way-too-high), so simply paste this into a bash-shell as root:
Code:
echo "manual" >/sys/class/drm/card3/device/power_dpm_force_performance_level
echo "vc 2 1801 1010" >/sys/class/drm/card3/device/pp_od_clk_voltage
echo "m 1 1150" >/sys/class/drm/card3/device/pp_od_clk_voltage
echo "s 1 1801" >/sys/class/drm/card3/device/pp_od_clk_voltage
echo "c" >/sys/class/drm/card3/device/pp_od_clk_voltage
/opt/rocm/bin/rocm-smi --setsclk 1 --setfan 90 -d 3
A few minutes later, after things settle down, rocm shows card 4 pulling 2w more, per-iter times hve dropped a smidge to 2318 us/iter for both jobs.

Gonna let things run like that overnight, if stable, tomorrow will upclock card 4 to sclk=2 and see what happens. But even at sclk=1, we have boosted our total system throughput by almost 75Miter @6M per day, the equivalent of one 6M-FFT PRP test every 36 hours or so.

Last fiddled with by ewmayer on 2020-08-11 at 22:21
ewmayer is offline   Reply With Quote
Old 2020-08-11, 23:15   #337
PhilF
 
PhilF's Avatar
 
"6800 descendent"
Feb 2005
Colorado

13258 Posts
Default

Quote:
Originally Posted by ewmayer View Post
[Nomenclatural note: henceforth will be using the initialism TDP to mean watss-at-wall-plug.]

when PSU manufacturers tout the wattage ratings of their gear, are they talking about watts of PSU *output*, or TDP?
It always means output. "TDP" will change, depending on the PSU's efficiency and supply voltage (240V would be more efficient than 120V).
PhilF is offline   Reply With Quote
Old 2020-08-11, 23:55   #338
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

11,743 Posts
Default

Quote:
Originally Posted by PhilF View Post
It always means output. "TDP" will change, depending on the PSU's efficiency and supply voltage (240V would be more efficient than 120V).
So for a reaonably surmise at PSU efficiency, based in part of the rocm wattages consumed by the cards - say between 80-90% - my current 900w TDP should put me safely within the PSU's 850w of rated ouput. And upping sclk to 2 romorrow - if the current single-Frankencable scheme allows it - will add ~50w, which should also still be OK. (He said, with his recently-purchased dry fire extingishers at the ready).
ewmayer is offline   Reply With Quote
Old 2020-08-12, 00:15   #339
chalsall
If I May
 
chalsall's Avatar
 
"Chris Halsall"
Sep 2002
Barbados

10,663 Posts
Default

Quote:
Originally Posted by ewmayer View Post
So for a reaonably surmise at PSU efficiency, based in part of the rocm wattages consumed by the cards - say between 80-90% - my current 900w TDP should put me safely within the PSU's 850w of rated ouput.
If you wanted to get really serious, you could measure the jueles of heat you needed to remove from the room...

I read a blog entry from the programmers of a Mars Rover a while ago, where they were calculating the costs of an AI process against an image in relation to their power budget (measured in joules).

Damn, those people are working at a different level than almost all other developers.

Last fiddled with by chalsall on 2020-08-12 at 00:18
chalsall is offline   Reply With Quote
Old 2020-08-12, 01:08   #340
PhilF
 
PhilF's Avatar
 
"6800 descendent"
Feb 2005
Colorado

52×29 Posts
Default

Quote:
Originally Posted by ewmayer View Post
So for a reaonably surmise at PSU efficiency, based in part of the rocm wattages consumed by the cards - say between 80-90% - my current 900w TDP should put me safely within the PSU's 850w of rated ouput. And upping sclk to 2 romorrow - if the current single-Frankencable scheme allows it - will add ~50w, which should also still be OK. (He said, with his recently-purchased dry fire extingishers at the ready).
You make sound like 849.5 watts is "safely within the PSU's 850w limit" lol.

If you like that power supply, and want it to live a long, happy life, you won't push it so close to it's limits.

This is becoming my favorite emote for you, lol:
PhilF is offline   Reply With Quote
Old 2020-08-12, 01:45   #341
sdbardwick
 
sdbardwick's Avatar
 
Aug 2002
North San Diego County

24×47 Posts
Default

Nah, you can run those Corsair RMs at 95% load 24/7 for at least 3 years at 35 degrees C; I'd say longer but mine has only been running for 38 months.

If his wattmeter is to be believed, he is currently drawing 783W DC, (900W x 87%+ efficiency at full load), or 92% of rated.
sdbardwick is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
AMD Radeon Pro WX 3200 ET_ GPU Computing 1 2019-07-04 11:02
Radeon Pro Vega II Duo (look at this monster) M344587487 GPU Computing 10 2019-06-18 14:00
What's the best project to run on a Radeon RX 480? jasong GPU Computing 0 2016-11-09 04:32
Radeon Pro Duo 0PolarBearsHere GPU Computing 0 2016-03-15 01:32
AMD Radeon R9 295X2 firejuggler GPU Computing 33 2014-09-03 21:42

All times are UTC. The time now is 21:15.


Thu Sep 29 21:15:57 UTC 2022 up 42 days, 18:44, 0 users, load averages: 1.43, 1.42, 1.35

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.

≠ ± ∓ ÷ × · − √ ‰ ⊗ ⊕ ⊖ ⊘ ⊙ ≤ ≥ ≦ ≧ ≨ ≩ ≺ ≻ ≼ ≽ ⊏ ⊐ ⊑ ⊒ ² ³ °
∠ ∟ ° ≅ ~ ‖ ⟂ ⫛
≡ ≜ ≈ ∝ ∞ ≪ ≫ ⌊⌋ ⌈⌉ ∘ ∏ ∐ ∑ ∧ ∨ ∩ ∪ ⨀ ⊕ ⊗ 𝖕 𝖖 𝖗 ⊲ ⊳
∅ ∖ ∁ ↦ ↣ ∩ ∪ ⊆ ⊂ ⊄ ⊊ ⊇ ⊃ ⊅ ⊋ ⊖ ∈ ∉ ∋ ∌ ℕ ℤ ℚ ℝ ℂ ℵ ℶ ℷ ℸ 𝓟
¬ ∨ ∧ ⊕ → ← ⇒ ⇐ ⇔ ∀ ∃ ∄ ∴ ∵ ⊤ ⊥ ⊢ ⊨ ⫤ ⊣ … ⋯ ⋮ ⋰ ⋱
∫ ∬ ∭ ∮ ∯ ∰ ∇ ∆ δ ∂ ℱ ℒ ℓ
𝛢𝛼 𝛣𝛽 𝛤𝛾 𝛥𝛿 𝛦𝜀𝜖 𝛧𝜁 𝛨𝜂 𝛩𝜃𝜗 𝛪𝜄 𝛫𝜅 𝛬𝜆 𝛭𝜇 𝛮𝜈 𝛯𝜉 𝛰𝜊 𝛱𝜋 𝛲𝜌 𝛴𝜎𝜍 𝛵𝜏 𝛶𝜐 𝛷𝜙𝜑 𝛸𝜒 𝛹𝜓 𝛺𝜔