Nvidia/AMD Cuda/OpenCL on Boinc projects - which card to buy?

Message boards : GPUs : Nvidia/AMD Cuda/OpenCL on Boinc projects - which card to buy?
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 · Next

AuthorMessage
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102455 - Posted: 6 Jan 2021, 18:14:43 UTC - in response to Message 102449.  

To true - I trust the properly calibrated instruments and loads that we used to determine what the commercially available PSUs were really capable of. Many (most) were just not worth the money as either they were incapable of sustaining their theoretical ratings or were somewhat inefficient (80% on a PSU is quite poor, the best designed commercial ones are running at over 90% across a wide range of loads).
ID: 102455 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102456 - Posted: 6 Jan 2021, 18:19:20 UTC - in response to Message 102454.  
Last modified: 6 Jan 2021, 18:26:02 UTC

If you are getting "luke warm" off a PSU then you either have a nice efficient one, or are not running at its rated load - remember a PSU of 1000w rated power and 80% efficiency will be dumping 250w somewhere and that will be a fair temperature rise given the poor cooling design of the normal domestic computer PSU.

[edit to add]
btw - Until recently I was using a Corsair 1000w (platinum) PSU, running at a real 900w output, the cooling air off that was at about ambient plus 40C - certainly not "luke warm" unless it was in a freezer, which it wasn't. The real input power was 1150w from our normal UK mains in the range 220 to 251v, 50Hz. Interesting was that the input power only moved about 3w over that input voltage range - output was solid at 900W and all rails sat nicely within their voltage tolerances, so the regulation was very good. The beast failed when I accidentally ramped to load to 1053w (smoke emerged).
ID: 102456 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102473 - Posted: 7 Jan 2021, 19:28:44 UTC

Well, you are clearly not running those power supplies at full power if you are only getting a "waft" of warm air. It sounds as if you are describing an ambient plus 20C.
The MEASURED efficiency of the half dozen 1kw Corsairs tested lay between 79 and 82% at full load, dropping to about 75% in the range 70 to 80% load. Which lines up pretty well with the figures published by Corsair.

That GPU that is tripping the PSU is probably drawing a very high start-up load. It's a shame you don't have a complete instrumentation set to see exactly what is happening. Your description of drawing a high, but far from excessive, current from a car battery does suggest a faulty on-board PSU device - or that it is actually trying to draw its rated power...... (20A at 12v is "only" 240w). (and I really "enjoy" replacing large surface mount devices - not)

How any switched mode PSU behaves on a given input voltage is very much down to how it is designed, some will be more efficient at 240V 50hz than at 110v 60hz and the same applies the other way round. Generally the difference is only a couple of percent, but I have seen over thirty percent loss of efficiency of 240v 50hz over 110v 60hz on one example form one manufacturer - I guess the iron losses were pretty bad on that one.
ID: 102473 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102488 - Posted: 8 Jan 2021, 18:40:36 UTC - in response to Message 102487.  

You need to take measurements and rely on your hand or guesswork.
I doubt that a GPU would drop the voltage on a "car battery" from 12v to 2v without the cables being in serious distress:
Did you get burns on your hands while doing this "test?
Did you actually measure the voltage?
Did you actually measure the current?

The first one is the only one that doesn't need instrumentation, and the only one where you really know the answer - The rest of the figures you have given are nothing more than PURE GUESS WORK and can thus be chucked in the bin as such,.

(Comment - if this GPU did this on a battery there is a fair probability that it would do much the same on any reasonable size PSU, after all "20A" at 12v is only 240W, and you keep saying you have PSUs, and 20A at 2v is 24w - you need to think and test properly before making such claims about the performance of PSUs.)

What "heating element" - you've just invented another part to your so called test.


You've read a lot "somewhere", and so have I, but I've done a whole load of designing and testing of power supplies over the years and KNOW how they work and how they behave, so give up now.
ID: 102488 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102494 - Posted: 9 Jan 2021, 14:40:40 UTC

Yes - I did get confused in all the diving between PSU & GPU.

Anyway, I've just caused an ancient and unreliable GPU to let its smoke out....
This GPU was removed from service because it would periodically resort to just displaying a series of broad bands and stop calculating followed shortly after with the computer crashing. It's been sitting on the self for some years. Dismantling it was a "work of art" - six different screw heads of three different patterns in use and a few self-destruct clips later I could see what the circuit looked like.
Starting from the power inputs. Those from the PCI connector and the Molex connectors are brought to one region of the board where they pass a number of voltage comparators comprising pairs of over & under voltage detectors. The truth line from each of those heads off into a single (unmarked) chip - presumably some sort of micro controller judging by its function (more later). The supply lines then head off to various other areas of the board - e.g. memory, fan, GPU chip and a couple of other locations I was too lazy to try and work out.
I had CMDs on all power input lines, and have access to a vast range of bench psu & logic simulators, plus a couple of 'scopes (and a lot of hook-up wires).
One of the output lines of the chip appeared to function as a "I've got Molex 12v" as "PCI has 5v & 12v". Having worked out what happened on the input I decided to skip the protection area for the Molex input (leaving the PCI connector set at its correct voltages), but the "12v Molex is OK" line forced high. Now the fun began - injecting power from a *-off variable power supply I raised the "12 v" line by a couple of volts, to no great effect, but when I started to drop the voltage things got a bit more exciting. At just over 5v the 12v to 3v DC-DC converter started to get a bit warm, dropping the supply voltage further at just under 4v smoke appeared from one of the power devices in the DC-DC converter and the current rose very rapidly from around 10A to well over 50A (limit of the CMD on the line from the bench psu in use), then the bench psu shut-down - trip set at a nominal 75A.
OK the voltages are different to those you saw, but this was not the same type of GPU, but I think it would be fair to assume that the power management is fairly similar across all types (at a functional block level, details like detection and trip levels will vary).

Well, what I saw lines up with what you describe. Neither of us can be sure that all GPUs have both over and under voltage detection on their supply inputs, but if it was missing, it is quite a simple step to see that having a very "soft" 12v line from the computer PSU could result in some quite spectacular destruction. I would hope that most consumer grade computer PSUs would have some form of under-voltage protection.
ID: 102494 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102499 - Posted: 9 Jan 2021, 19:08:06 UTC - in response to Message 102497.  

I wasn't looking to fix it - I was looking to see what happened at low supply voltages - I sort of hoped the 12v to 3v dc-dc converters would be sufficiently well designed to shut down if the voltage dropped, but they were controlled by the micro-controller, which of course I'd cut out of the equation.
No shame on you - GPUs are not really designed to be user serviceable to any great extent.

Your 11.5v supply is very close to the trip voltage of the under-volt detection on the GPU I destroyed (11.3v), so given the sort of tolerances these things have suggests you are just on the "safe" side of the line.

Those SM capacitors aren't there to filter noise, they are there to decouple the various supply rails. There are probably a number of fairly large (10mm sort of dimension) that are there to do supply-rail smoothing. On the GPU I had these were fixed pin, through-hole devices, the way the leads had been cropped which made them look as if they were SM on the back of the board, but very careful study showed their true colours.
ID: 102499 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102512 - Posted: 10 Jan 2021, 19:11:59 UTC - in response to Message 102510.  

Wikipedia has over simplified the function of a decoupling vs. noise filtering. Decoupling is used to stop stray signals, or DC, from crossing between circuits. Noise filtering is used to block the noise normally associated with switching circuits, clocks etc. during normal operation. A subtle difference, and I've oversimplified things somewhat.....
ID: 102512 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102513 - Posted: 10 Jan 2021, 19:17:29 UTC - in response to Message 102510.  

If they released circuit diagrams it would make it easier. Not really easy to trace circuits when they're so small, go under things, and have multilayer boards.


Many complex consumer-level devices have been the subject of many man-years of design, development, testing and the manufactures want to protect this investment so they don't publish their PCB designs (which almost certainly includes more than just the schematic, but "interesting" things like screening layers, power layers. I agree it's a shame, but they have their reasons, we can either like them or not......
ID: 102513 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102528 - Posted: 11 Jan 2021, 19:22:30 UTC - in response to Message 102527.  

Yes, but how many tens of pages do you want - I suggest you dig around and find a copy of "The Art of Electronics" by Horowitz & Hill as that will save me typing. There are any number of other electronics text books, but few are quite as simple to understand and few give so many good, workable examples.

The sort of power supplies used in many small appliance battery chargers are switched mode, and rely on either RC or RL filters to get rid of the residual high frequency that is used to produce the low DC voltage. In other types called "capacitive dropper" there is an RC network, combined with a simple bridge rectifier drops the mains voltage to that required to charge the battery - but those tend only to be used on very low-power batteries.

As I implied a few posts ago, the design of PSUs is not just a simple matter of slinging a few components together, but these days is much more involved - ah the days of the nice simple capacitive dropper used for the heater circuits on valve radios.....
ID: 102528 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 102530 - Posted: 11 Jan 2021, 19:48:12 UTC - in response to Message 102529.  
Last modified: 11 Jan 2021, 19:50:27 UTC

Five sentences would lead you into more confusion, just do as I suggested and you won't be confused. This is one of those subjects where there is nothing really of use between the very simple explanation I gave earlier and diving into the theory.

Capacitive droppers have, by dint of the way they work extremely poor power factors, and are at anything over low loads very inefficient. Linear & switched mode psu should have much better power factors, and thus better efficiency (but if poorly designed or implemented can have very poor efficiency and/or power factors).

edit to add
Capacitive droppers are not switched mode supplies, the latter run at high frequencies, where as capacitive droppers run at the frequency of the incoming supply.
ID: 102530 · Report as offensive
Previous · 1 · 2 · 3 · 4 · 5 · 6 · Next

Message boards : GPUs : Nvidia/AMD Cuda/OpenCL on Boinc projects - which card to buy?

Copyright © 2024 University of California.
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.