Full dies always overclock better than a Cut down.

Empirically this is correct, but with everything else being equal a full die would clock worse simply because it produces more heat. And the cut-down version can have the slowest part disabled, which can yield more frequency headroom.

MrS

It is a bit of a puzzle why the cut-down chips don't do better. I expect it is because they disable portions of the active circuitry that are not functioning correctly, but leave behind the clock lines. That just maintains the capacitive load without the means to drive it at full speed.

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-960,4038-8.html

Includes PCIe/PEG measurements and power targets (120-160W) for most GTX960. Evga provides a 8 pin while Gigabyte G1 is [2] 6pin and remaining boards are [1] 6pin. Looking at guru3d thermal shots: Galax has the coolest VRM and core temps.
Newegg lists GTX960 at 199-209usd.

It is a bit of a puzzle why the cut-down chips don't do better. I expect it is because they disable portions of the active circuitry that are not functioning correctly, but leave behind the clock lines. That just maintains the capacitive load without the means to drive it at full speed.

Well, I'm fine with my explanation given in my last post.

The point you're making here has one weak spot: we're talking about lot's of millions transistors here. If the capacitance of fused-off transistors would hamper the performance of transistors in a neighbouring SM, we'd have a huge capacitive problem. At usual transistor densities (as close as possible) the capacitive load would be prohibitively high [if this was true] and our entire chip design, scaling and technology development would have to be changed. Luckily it's not that bad :)

On topic: it should be interesting how GTX960 actually performs here. 8/5 or 60% better than a GTX750Ti is expected, but the memory speed doesn't scale as well as the crunching power (30% higher bandwidth due to clock speed). This difference doesn't sound dramatic, so we may well see performance in the range of 50 - 60% higher.

MrS

---

Scanning for our furry friends since Jan 2002

The point you're making here has one weak spot: we're talking about lot's of millions transistors here. If the capacitance of fused-off transistors would hamper the performance of transistors in a neighbouring SM, we'd have a huge capacitive problem. At usual transistor densities (as close as possible) the capacitive load would be prohibitively high [if this was true] and our entire chip design, scaling and technology development would have to be changed. Luckily it's not that bad :)

No, I am referring to the "clock lines", which are the metallic conductors that carry the clock signals over the chip. They would not be so easy to disconnect when you do a chip repair; it is easier just to turn off transistors, and so the clock lines might be left in place. They would then be a large load on the clock-drivers that are left operational. That is a bit of speculation of course; it could have to do with various other portions of the circuitry, but it appears to be something basic. Nvidia would not want to lose performance on the cut-down chips if they didn't have to.

Ah, now I got your point. Power consumption of the clock signal is indeed a serious issue in modern chips. But during recent years there has often been talk about better "clock gating" when new chips were presented. I always understood this as not delivering the clock signal to regions of the chips which are currently not in use, i.e. power gated. If the ycan do this, they can also clock gate deactivated SMMs.

MrS

---

Scanning for our furry friends since Jan 2002

I always understood this as not delivering the clock signal to regions of the chips which are currently not in use, i.e. power gated. If the ycan do this, they can also clock gate deactivated SMMs.

Possibly so. But there are clock lines and then there are clock lines. Some are "local", which would be easier to turn off, and some are "global", which might not be. And some would be intermediate between the two. What you do for repair is probably different than what you do in normal operation, but beyond that is beyond the scope of this discussion I am sure.

My GTX 980 was getting to noisy for me, so I swapped it out for a EVGA GTX 960 SSC (mainly to test the ACX2.0+ cooler). The money I've saved will probably be used for a Titan 2 or I might as well just stick with the 960 until Pascal comes out.

I still have to monitor GPUGrid performance. But I don't expect much. Power draw has decreased by 45W (-15%). Ok, I guess. But only if it doesn't fail crunching :/ The card is dead silent. Fans are spinning with 670 rpm under full load (Noelia), temperature never goes beyond 67°C. CPU temp has gone up by 1-3°C due to to cooler not exhausting heat out of my µATX case. But it's barely noticeable.

They only thing I can hear know while crunching are my 3 Noctua 120 mm fans spinning at max. 850 rpm (CPU fan). It's really quient now. No comparison to the blower with it 2000+ rpm and 80°C. Fun fact is, that the reference 980 got quite toasty on its surface and heated up other components around the GPU as well.

I did this mainly to test if my case could handle a non reference card.

Quickly looking at your results shows the GTX960 card being a bit faster than half the performance of your GTX980 - which is nice, since it's only got half the raw power. But saving only 45 W? Given the TDPs this number is plausible, but is that worth it? You could have simply lowered the power target on the GTX980 or lowered the fan speed, which would have made it boost less while staying at 80°C.

Anyway, I'm sure you got some good money for your GTX980 and if you're happy with the GTX960, so be it :)

MrS

---

Scanning for our furry friends since Jan 2002

Hi Extra,

my GTX 980 power target was already lowered quite significantly (around 74%) so that it stayed at around 1240MHz (dropping down to 1216MHz occasionally). I probably could have taken it furhter than this, but I didn't want to get too far away from stock clocks.

My GTX 960 is now tweaked the same way (even further, I've lowerered it to 55% power target (!), voltage is at 1,1V-1,13V, clocks stay at around 1340Mhz.

After taking four Noelia WUs into account, I can now say:
GTX 980, AVG: 6,4 hrs
GTX 960, AVG: 10,7 hrs

Hm, dunno if my math is right. But this seems pretty disappointing. It takes 70% more time now to complete a WU, yet I only save around 15% power. So actual efficieny is quite low, isn't it?

What does matter however is that the card is now very silent even at full load and less heat is exhausted into my case (yep, you heard me right, less heat, even though this is not a DHE design).

Saving money for big Maxwell now :) This card will eventuelly go into my dad's PC.

Thanks for the more precise numbers, Dave!

They still look good: your GTX960 has somewhat higher clock speeds and average GPU utilization is generally higher for smaller cards (which explains the remaining performance advantage of the GTX960).

Be careful when judging power draw, though: many cards have power targets set in their BIOs which differ from nVidias recommended values. You can not easily see this, as all you're given are percentages. "MaxwellBiosTweaker" can read it out (and modify it, if one wants to). Your GTX960 should currently draw about 100 W, whereas a typical power target for GTX980 cards is 180 W. A card with blower-style cooler has probably no higher setting than this, but it may still be more than nVidias stock value of 165 W. So I estimate you're saving 60 - 80 W or % now, which is fine considering the performance. Hard measurements would be better, for sure, but you'd need a power meter for this.

Edit: or look into your GPU BIOS, so you know exactly which power draw xx% translates into.

MrS

---

Scanning for our furry friends since Jan 2002

The only valid way to measure power draw is to measure it. Use a power meter such as the very affordable Kill-a-watt or equivalent. The theories are fine but until you measure it you just don't know.

Huh? I did measure power draw.

It went from 320W down to 275W. Approx. -15% compared to my GTX 980. Both with lower power limit.

I know, was referring to all the theory going on elsewhere. Sorry for the misunderstanding, I should have been more clear. Theorizing is fine to a point but actual measurement such as you did is the only real way to know. I've been fooled more than once by assuming things about power draw and then finding out I was all wet after measuring.

You appear to be crunching Einstein on your CPU.
Ask yourself a question - Why?

Huh? I did measure power draw.

It went from 320W down to 275W. Approx. -15% compared to my GTX 980. Both with lower power limit.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

You appear to be crunching Einstein on your CPU.
Ask yourself a question - Why?

Huh? I did measure power draw.

It went from 320W down to 275W. Approx. -15% compared to my GTX 980. Both with lower power limit.

Well, universe, space. Look at the stars at a clear night ;)

Funny you asked though. Because in fact I stopped crunching E@H and switched to Rosetta@Home for my CPU. R@H seems to need all the CPU power it can get.

You appear to be crunching Einstein on your CPU.
Ask yourself a question - Why?

Huh? I did measure power draw.
It went from 320W down to 275W. Approx. -15% compared to my GTX 980. Both with lower power limit.

Well, universe, space. Look at the stars at a clear night ;)

Funny you asked though. Because in fact I stopped crunching E@H and switched to Rosetta@Home for my CPU. R@H seems to need all the CPU power it can get.

I think the subject of skgiven's question is not your motivation for crunching Einstein@home, but the reason for the power drop you've measured.
That reason is that the other project's applications could not utilize the latest GPUs (regardless of any GPU utilization readings by different tools) as much as GPUGrid does, partly because the other projects are using older CUDA versions.

Oh, no, the power drop is real. Everything else stayed the same. All I did was swapping out the GTX 980 for a 960. The CPU utilization remained unchanged.
Some minor misunderstanding going on here I guess haha but thanks anyway for the input.

The power drop from 320W to 275W was simply explained by the GPU TDP/usage drop; 165 to 120 is 45W.
With respect to the GPU's this is a 37.5% drop but to the systems power usage it's a ~15% drop (275/320=0.859).
A 15% reduction in system power usage vs a >40% performance loss isn't good.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Every time I peak at my 960 it's doing a short run, is it just me?

---

Every time I peak at my 960 it's doing a short run, is it just me?

It's probably set in your preferences that you accept work only from the short queue (perhaps it's the default setting and you haven't changed it).