Inconclusive results

Started by Caprio, June 05, 2012, 04:09:47 PM

Previous topic - Next topic
Hello,

I ran 3 tests of optimized clients for GPU 560Ti.
Card B has BIOS version 70.24.21.00.02 (P1040-0050)
Card C has BIOS version 70.24.2E.00.02 (P1040-0050)

Card B was purchased before Card C. I have experienced more inconclusive results with Card C.

Let me Post the WU data separately.

Here is

1) the WU, for which I ran the test.
2) Bench configuration file

The driver version was 301.42.

One thing I noticed was that Knabench is able to pull out a GPU from 405MHz locked state.

June 05, 2012, 06:23:44 PM #2 Last Edit: June 05, 2012, 06:39:49 PM by Jason G
Thanks for the Detailed tests.

  card C is either outright broken, or at least needs more Core voltage than it has (alternatively frequencies backed off a lot).

Card B is also 'on the edge' of stability, but not stable enough.

Best bet is up the core voltages on boths cards to ~1.062V  .  ASUS factory 1.025V, is not enough for these GPUs to do 900MHz stable Cuda crunching, period.  Default is 1.05-1.068V for most other brands at that frequency, with that same silicon.

Factory undervolt by some brands/cards is a known problem, particularly with early superclocked 560ti.

You've previously informed the PSU is Corsair 650W. That will run 1 of these cards installed only at a time.  With either one of those cards installed, rerunning the benches with core voltage raised should verify matching results. 

Yes, multiple issues at play can be tricky to sort out, but these are definites from experience since 560ti came out.
Card B's GPU silicon is slightly lower leakage, so tolerates the undervolts slightly better (but not good enough).

HTH

Jason


It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change.
Charles Darwin
---
Chaos: When the present determines the future, but the approximate present does not approximately determine the future.
Edward Lorenz

Maybe it is best to RMA the card C as it has behaved the same from start.

June 06, 2012, 07:00:05 AM #4 Last Edit: June 06, 2012, 07:55:09 AM by Jason G
Quote from: Caprio on June 06, 2012, 01:52:27 AM
Maybe it is best to RMA the card C as it has behaved the same from start.

Personally, if It was me and I didn't want to boost voltage to attain stable crunching at factory frequencies, or lower clocks, I would *try* RMA both cards.

The complication with RMA is, depending on the warranty, you may get the same cards back, as 'not defective for games' as they are designed, or,  replacements with the same too low voltage back.

That's the problem some brands made by releasing superclocked cards with too low default voltage. (Well it isn't too low for games, games can tolerate some artefacts, and don't use the whole chip)

Jason
It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change.
Charles Darwin
---
Chaos: When the present determines the future, but the approximate present does not approximately determine the future.
Edward Lorenz

No matter what you decide, definitely up the voltage on those cards. MSI Afterburner will allow you to do it with no problem.

http://event.msi.com/vga/afterburner/download.htm

It is best to at least try RMA the card. Once the manufacturer in this case ASUS starts getting back the cards, what I hope is that they stop putting "unstable" cards to the market with too low voltage. ASUS got greedy on being green. They should use KnaBench to test their cards in production for stability and use such a voltage that will give a successful yield rate of 99%.

Actually the card should start with a safe high voltage in their defalut BIOS production. Then it is up to each customer if they want to undervolt the card with some tweaks.

I have even another card same ASUS 560Ti TOP in my other host. This card actually runs on higher voltage 1.05V. It has not given one inconclusive result.

What I would say to ASUS:"You released the card with certain working frequency (GHZ) and Voltage. You are held accountable for your design. For example issue a new BIOS in which all cards are then operating with 1.05V instead of 1.025V if it makes the difference in stability.

BTW, NVIDIA proposed that I should try out the 280.26 driver (for the downclock issue, which occurs in Starcraft 2). The effect is that even the standard menu screens runs at 4fps.

I also rerun the testcase with new driver. To my suprise the execution is faster with 280.26 driver compared to 301.42.

Rgs,
Caprio



Card A has same BIOS as Card B. Why is there a difference in the reported Voltage?


June 07, 2012, 02:21:59 PM #8 Last Edit: June 07, 2012, 02:39:36 PM by Jason G
Quote from: Caprio on June 07, 2012, 06:18:27 AM
Card A has same BIOS as Card B. Why is there a difference in the reported Voltage?

Only the manufacturer could give you a definite answer on that, but my guess is that the BIOS or card itself has some capability for tuning adjustment of the voltage, using the same BIOS revision.  If something like that was available at the time of manufacture, they would have guaged the 'correct' voltages, (which we know are really too low), by some tests on the GPU (similar to OCCT or  Furmark etc) then set the BIOS clocks & voltages accordingly for acceptable artefacts.... Then labelled the card as superclocked (or whatever) pidgeonhole it reached.  Some companies are better at this 'parts binning' process than others, which is part of quality control, and that applies in every electronic industry area.

Every GPU piece of silicon is a bit different, made using many layers including processes like 'Chemical Vapour Deposition', which is a fancy-hitech name for spraypainting. There is substantial variation on nanometre scales (as well as defects), which means different GPUs have different leakage properties, as so require different voltage to reach a specified power or performance criteria. 

1.025V is of course, too low for any of these GPUs to crunch at 900MHz stably, nomatter what the leakage specs are.  Probably on games you won't notice the visual artefacts they deemed acceptable.

When these cards came out, they were competing across brands for the huge price/performance mid-range market, which is very competitive.  nVidia GPUs are a very expensive component, so it shouldn't be a surprise that some manufacturers might be pushing the batches of silicon they receive a little harder than others, so they stay below power & thermal design envelopes at a raised frequency.  Unfortunately, as with overclocking a card yourself, this takes more care & attention to detail than some production environments budget for, so while the GPU itself may be 'fine' they sometimes don't test long enough, or under the strictest conditions, as we would at home over several hours, so preset too low a voltage. 

QuoteI also rerun the testcase with new driver. To my suprise the execution is faster with 280.26 driver compared to 301.42.

x41g is a Cuda 3.2 build, so it's logical that under some conditions the 280 series Cuda 3.2 driver would perform better.  These differences tend to iron out over full length runs.  301.42 is a Cuda 4.2 driver, and Cuda 4.1 & Cuda 4.2 builds are superior on these GPUs... so if & when you can get it stable for starcraft on a newer driver, the card can do a lot more yet.  (The downclock with stracraft is most likely also stability/voltage related, a driver crash induced by similar starving... They have suggested you use the old driver to remove the protection mechanism... better IMO to actually fix the problem, which is too low core voltage)

Jason
It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change.
Charles Darwin
---
Chaos: When the present determines the future, but the approximate present does not approximately determine the future.
Edward Lorenz

I returned the 560Ti to the seller. The importer was not happy to accept a return. However the seller is a esteemed place, which puts customer first and was happy to give me a full refund.

Meanwhile I have run Card B with 1.037V and 1.05V. The results are all the same for Card B as with 1.025V.

Now I put my eyes on some other card. Maybe 6x0 lineup. I will steer clear of any "factory" superclocked cards. More cores the better. Performance / buck / Wattage is what im looking for.




June 08, 2012, 11:44:48 AM #10 Last Edit: June 08, 2012, 11:49:35 AM by Jason G
Quote from: Caprio on June 08, 2012, 07:41:59 AMNow I put my eyes on some other card. Maybe 6x0 lineup. I will steer clear of any "factory" superclocked cards. More cores the better. Performance / buck / Wattage is what im looking for.

As one option to consider, If the initial cost can be managed, a reference 680 cleans up in the performance/buck/Wattage arena with x41x, over all other single GPUs (for the moment).
- Initial cost High  (I stretched the budget a lot, though have found it's been well worth it)
- Ongoing Power cost low (a bit less usage than superclocked 560ti)
- Highest Cuda performance for single GPU (at the moment) running x41x
- self stabilising/boosting in hardware, voltage + modest OC, so previous issue as with factory SC 560ti's is gone.
- low heat & noise
- with software & drivers, at the start of the optimisation curve, whereas 560ti would be in middle toward mature
- Will run any game I've tried (not that many) on max settings at 1080p resolutions,

The way I see it, they engineered everything out of this card that could be seen as mistakes in earlier generations, so it'd be the way to go if you want something to work 'out of the box'.  The 560ti was always an 'enthusiast' card, requiring extra tweaking & care, targeted into the wrong market segment IMO, whereas the 680 was designed to be 'plug n' play'.

Jason
It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change.
Charles Darwin
---
Chaos: When the present determines the future, but the approximate present does not approximately determine the future.
Edward Lorenz

Powered by EzPortal