How come the GTX 1070/80 FE cards seem to be just as potent overclockers as the AIBs?

[hsjj3]

So I'm curious about this.

Looking back at previous Nvidia generations, a reference model card was never as good an overclocker as the AIB cards. More than anything else, the power limit seemed to have been the limiting factor.

Now though, I'm seeing overclocking results, and there is absolutely nothing to suggest that the AIB cards with extra power connectors and higher power limits can overclock particularly any better than the reference FE models.

To be clear, I refer to "normal" overclocking for both the above cases using consumer tools such as MSI Afterburner, meaning I do not refer to BIOS edits that increase voltage levels. That was always "higher" overclocking that a small segment of the community dabbled in.

 

[Flapdrol1337]

Nvidia demands very stringent limits on power and voltage.

If a gpu breaks the aib partner sends it back to nvidia for warranty. By having these limits it's near impossible for a user to break their gpu.

 

[DisarmedDespot]

IIRC the chips themselves have a hard voltage limit that can't be bypassed. Basically, even if AIB cards have the power delivery and cooling to safely handle more voltage, the chip physically won't take it.

 

[fastamdman]

There are some voltage mods you can do to these cards as well. I don't know how aware people are of this, but I will toss the link in here.

https://xdevs.com/guide/pascal_oc/

 

[Flapdrol1337]

IIRC the chips themselves have a hard voltage limit that can't be bypassed. Basically, even if AIB cards have the power delivery and cooling to safely handle more voltage, the chip physically won't take it.

This 1.25V "hard limit" really isn't the issue. Without warranty breaking mods they won't even do 1.1V

 

[Kenmitch]

My Zotac 1070 AMP was delivered last night pretty late. Didn't get much time to play around as the 144Hz DVI bug got me when drivers installed....Black screened. Switched to HDMI till amazon delivers my DP cable today.

Default clocks which boosted @1999MHz in my limited testing so far. Didn't try OC yet as I'm waiting for my DP cable. Not expecting any crazy overclocks as default pretty much slaps my CF 390's around @1080p anyways. Whisper quiet as I only hear my H100i and case fans.

 

[Qwertilot]

Also perhaps fair to admit that NV did do a decent job on the FE cards, as opposed to a fair few historical reference cards. Not justifying the premium of course

 

[exar333]

Similar to 670/680 where the cards were more freq-limited vs. thermal limited.

The 1080/1070s are (for the most part) hitting a wall around 2100mhz. Better power delivery, etc. doesn't seem to be helping much to go over that substantially.

 

[alcoholbob]

It only seems like the cards have less overclocking headroom because Nvidia increased it's midrange TDP and base clockspeeds this generation on its reference cards in order to make their cards look faster. But all they did was cut into the overclocking headroom that they had left to users in the Kepler and Maxwell generations.

 

[exar333]

It only seems like the cards have less overclocking headroom because Nvidia increased it's midrange TDP and base clockspeeds this generation on its reference cards in order to make their cards look faster. But all they did was cut into the overclocking headroom that they had left to users in the Kepler and Maxwell generations.

They are slightly higher than Maxwell, and lower than Kepler in TDP for mid-range. I would say they are pretty much on-par.

You could be right about the clock-speed limits, but the arch feels a lot like 6xx Kepler where the cards just didn't go over a specific speed. I run just over 2050mhz constant, with low temps and no extra voltage. If the chip allowed a higher speed, I am very confident I could get it to go a decent amount higher. It just doesn't though...

 

[realibrad]

There are a few reasons that Nvidia might want to limit OC.

A while back, Nvidia got hit with a lawsuit for cards failing. By limiting OC headroom, the cards should last longer. Nvidia leading the market has no reason to take the risk again.

2nd, this node and all smaller nodes will not OC like before. Frequencies are going to be much tighter. We are starting to run into issues like quantum tunneling because we have pushed the current materials to the limit. Adding more voltage would only make things more unstable.

3rd goes back to market share again. Nvidia having something like 80% of the market means they dont want to let people OC when they can release another card and get people to buy those. A 1080 that could OC higher could mean the next card would not be seen as worth an upgrade.

 

[littleg]

I don't think it's intentionally limited, I think they've pushed it right up close to the limit to achieve the performance they were targeting. Maxwell had bags of headroom because they could achieve their performance targets without needing to pump the clocks as high as they could. That let the cards use less power while still getting them to where they wanted to be.

With Pascal I think they have had to push the clocks higher than maybe they would have otherwise liked to hit their performance envelope. This naturally means less headroom for overclocking.

 

[Keysplayr]

I don't think it's intentionally limited, I think they've pushed it right up close to the limit to achieve the performance they were targeting. Maxwell had bags of headroom because they could achieve their performance targets without needing to pump the clocks as high as they could. That let the cards use less power while still getting them to where they wanted to be.

With Pascal I think they have had to push the clocks higher than maybe they would have otherwise liked to hit their performance envelope. This naturally means less headroom for overclocking.

Also if they pump up the clocks, they can afford to use less transistors for smaller dies. As you can see, a 2560 cuda core 1080 can outpace a 3072 cuda core TitanX. Less Raster units, less everything. If the clocks were limited to around 1200MHz or so, they would need more cc and more die real estate. Saves money all around.