Prime95 performance suffers after installing GT 640

[Rodrigo]

Originally Posted by LaurV

This related to power supplies...
An old man goes to the doctor, accusing a strong pain in his left leg. "It is because of aging, Dad.", the doctor said trying to calm him down, "yeah, but you see", answered the old man, "the right leg has the same age, and it does not hurt..."

LOL, I do feel this is the sort of thing that's going on.

See the next post...

Rodrigo

[Rodrigo]

New installment in the soap opera, "As the GPU Fan Turns":

Late last night, before going to bed I decided to take out the new 630 I'd put in there so that, at least, the dx7500 would be LL'ing at the faster rate.

When I wiggled the mouse to wake up the monitor, before powering off I had to stop Prime95. Imagine my surprise when I saw that the per-iteration times, with the GT 630 still in place, were now at 66 and 71 ms. max:

All I'd done earlier that night was to stop the program in order to get a wattage reading, and then to start it back up. (I didn't even close it out, just used the Stop/Continue function.)

Now how does this affect the diagnosis?

[Rodrigo]

Originally Posted by Antonio

I have been following this thread and would point out the following.
1. Both the 630 and the 640 are powered only from the PCIe slot so the maximum power draw is 75W, this can consist of (at maximum) 10W from the 3.3V supply and 65W from the 12V supply. So it looks like nVidia are just quoting the PCIe spec!

2. I have a Zotac GT640, and from measurements I have made this card draws 8W on idle and 48W when running mfaktc (0.20), and has no effect at all on the timings of 2*(P-1) and 1*LL tests running on Prime95. In fact I have the GT640 and a GTX650Ti in the same computer, both running mfaktc without affecting Prime95 timings.

3. Considering the sequence of events outlined in the thread I would imagine that the GT640 was in fact faulty from the beginning and has now effectively died. Possibly a bad (or reverse polarity) capacitor in the voltage convertor circuit. Continuing to try to use it could damage the motherboard and/or the power supply. Trying it with a more powerful supply could increase the possibility of damaging the motherboard as more amps would be available to burn out the PCIe power tracks.

Thank you Antonio, this makes sense. I'll pack up the 640 as chalsall suggests and see if I can send it back.

Rodrigo

[Rodrigo]

Originally Posted by chalsall

(And, to put on the table, I made a bit of a mistake in my counsel above. When you said the box mentioned 20A, I assumed that's what the card drew. Drilling down further, however, it appears that's what the minimum PSU should be -- the card draws much less.)

So then, if the card draws much less, do you think we should be OK with the existing PSU and the 630?

Rodrigo

[Aramis Wyler]

This may be a little seat of the pants, but it sounds a bit like a resource manager is limiting the usage of some applications based on the usage of others. I don't recall every hearing of such a thing wrt power, but who knows, windows is a many layered beast. At the end of the day though, that would still mean you were pushing the envelope for power. It's clear your power supply is not quite meeting the specs for what you're doing with it, and this sort of ghostly phenomena is not entirely unexpected when you're in that sortof - not technically good enough but maybe if I'm lucky it will work - state.

PS. Most power supplys operate at peak efficency when they are supplying half of their max output. The efficiency of a power supply is a curve, not a straight line.

[chalsall]

Originally Posted by Rodrigo

So then, if the card draws much less, do you think we should be OK with the existing PSU and the 630?

The supplier wants you to have a 20A PSU. You have a 19A. It is common in this industry to overspec the requirements for power. And since the type of work we do actually taxes the GPUs less than games do, you're probably OK.

[Rodrigo]

Originally Posted by Aramis Wyler

This may be a little seat of the pants, but it sounds a bit like a resource manager is limiting the usage of some applications based on the usage of others. I don't recall every hearing of such a thing wrt power, but who knows, windows is a many layered beast. At the end of the day though, that would still mean you were pushing the envelope for power. It's clear your power supply is not quite meeting the specs for what you're doing with it, and this sort of ghostly phenomena is not entirely unexpected when you're in that sortof - not technically good enough but maybe if I'm lucky it will work - state.

PS. Most power supplys operate at peak efficency when they are supplying half of their max output. The efficiency of a power supply is a curve, not a straight line.

Thanks for that info!

One (more) odd thing about this whole series of events is that sometimes when I restart either of the dx7500's and relaunch Prime95, the program operates at a significantly slower rate, but then a reboot usually takes care of that. This was happening on the "bad" dx7500 even before I put in any kind of video card. The difference now (with the card installed) was that the poor performance persisted through several reboots. And now it's working fine. Go figure.

Rodrigo

[Rodrigo]

Originally Posted by chalsall

The supplier wants you to have a 20A PSU. You have a 19A. It is common in this industry to overspec the requirements for power. And since the type of work we do actually taxes the GPUs less than games do, you're probably OK.

All in all, that's a relief -- thank you!

Rodrigo

[Xyzzy]

Maybe try to eliminate the OS variable?

You could boot from a Linux "Live CD" or USB key and test with the same exponents.

[Rodrigo]

Originally Posted by Xyzzy

Maybe try to eliminate the OS variable?

You could boot from a Linux "Live CD" or USB key and test with the same exponents.

If it helps in determining the answer to this, I dual-boot the other ("good") dx7500 with XP and Vista Business, and the LL per-iteration times are comparable to each other.

But I will definitely try your suggestion if the current LL times become consistent across multiple reboots. Otherwise, whatever readings we get we can't be sure that they're due to using a different OS or to the sort of maddening capriciousness we've already seen!

Rodrigo