Why don’t they make a low profile gpu with a powerful mobile GPU?

[tkitch]

Realistically I don't think you're going to get something much more powerful than a 1650 Super at this point, without additional power.

 

About the only thing adding RTX cores would get you, would be RTX for Screenshots.  The 2060 can't run games with RTX enabled, it just doens't have the power.  

 

You can't rewrite physics to get something you would like, currently.

[Mark Kaine]

12 hours ago, porina said:

If anyone has an Ampere card

So I undervolted my card a little, it's harder than I thought to find the right balance - especially because I had Shadowplay running also - which I've noticed later makes the card boost like crazy even with the undervolt... 

 

 

However once I turned this off things got a bit more straight forward.

 

The lowest (roughly) I got was 0.756v , at 80% usage, 1400mhz 'boost', 50C, 1440p, 60fps... I didn't have gpuz running otherwise I would probably know the max wattage (although that's not necessarily representative, due to spikes and whatnot) 

 

But can't you or someone else do the math now - 756mv/ wattage ?

 

that's barely above idle voltage btw, and I had memory at +500 also. 

[porina]

31 minutes ago, Mark Kaine said:

But can't you or someone else do the math now - 756mv/ wattage ?. 

Too many variables, which either need observing or prior knowledge.

 

I think I'll have a go on Turing for illustration. Spreadsheet on standby! At most I'd extrapolate within a generation. I would not want to try extrapolating between different generations.

[Mark Kaine]

4 minutes ago, porina said:

Too many variables, which either need observing or prior knowledge.

 

I think I'll have a go on Turing for illustration. Spreadsheet on standby! At most I'd extrapolate within a generation. I would not want to try extrapolating between different generations.

yeah i thought so, well idk I thought maybe voltage would be enough. 

 

but yeah Im trying with gpuz now but I'm already seeing despite 'undervoltage' it already spiked once to 800mv... so that's something you'd need the card to not do anymore for a start, that'd probably fry any pcie slot - but yeah I believe with this setting it should be roughly 150-170w "board power draw" (of course it's not a mobile card, so) 

[Mark Kaine]

so for the record with undervolt @ ~700-800mv board power draw is around 120-165w... at no noticeable performance cost in my testing scenario. 

 

 

which is pretty low but not 75w low... I still think at some point its not worth it anymore and not a 3070 anymore either ... but I think a low power ampere 'mini' could probably be done (at what performance is still the question) 

[porina]

Did some testing but didn't have as much adjustment range as I thought I might have so it will be limited in scope.

 

System is 2070 +8086k (turbo off). I adjusted the power limit on the 2070 and observed the changes to performance and power draw. Test Unigene Heaven, 1080p extreme, manually moved view in front of the dragon statue and left there so it was consistent. I left the test running for a bit after making a change as thermals do impact performance also. Only then I took an average.

 

Main settings I'll describe will be: stock (100%), max power limit (114%), min power limit (71%), and min power limit + mem speed reduced as far as I can in MSI afterburner (-502).

 

Increasing the power limit from 100% to 114% increased board power from 173W to 189W (+9%), with no change to fps at 83 for both. GPU clock hardly changed from 1785 to 1800. Mem clock didn't budge. Basically it just makes the card angrier without getting anywhere.

 

Reducing power limit to minimum of 71% (compared to 100%) did make a difference. Board power went down to 124W (-28%) with fps drop to 76 (-8%). Core clock was now 1637 (-8%) which matches the fps drop.

 

At 71% power limit, reducing the ram clocks didn't change board power for overall fps in this example. Ram clock dropped from 1700 to 1575 (-7%). Whatever savings were present from reducing ram speed went into increasing core power which increased fractionally from 1637 to 1647. Reported memory power difference was within 1-2W so I'm not confident in calling this significant. Also, if you add the reported vram powers to the core power, the number significantly exceeds the reported board power. So something isn't correct there. If you add up the reported PCIe slot and power connector powers together, it does equal the total board power within tolerance. So I'd take the core/vram power values to be less trustworthy without understanding them better.

 

I also noted the reported memory power usage in these and other usage scenarios. The power usage of vram did reduce slightly as I reduced the power limit, but I'm hesitant to reach a conclusion here as there wasn't a strong effect.

 

Given the above I'm not confident in extrapolating how low you might have to go to get anywhere near 75W TBP. We clearly see the cards running into the less efficient zone under normal operating conditions, given that a 28% reduction in power limit only resulted in a 8% drop in fps, at least for this particular test scenario. How much further that curve goes I can't test far enough down to find out. I'm sure I had other nvidia GPUs which allowed much lower power limit than the 71% of this particular sample.

 

Just thought, I probably should have done some observation to eliminate the possibility of CPU limiting... will get back on that.

[Mark Kaine]

23 minutes ago, porina said:

Just thought, I probably should have done some observation to eliminate the possibility of CPU limiting... will get back on that.

Well you also should try a custom voltage / frequency curve because that reduces power draw significantly compared to just setting a power limit in my testing... 

 

I kinda think keeping the voltage low is key here, but seeing im almost at idle voltage already idk how much further it can go.

[porina]

35 minutes ago, porina said:

Just thought, I probably should have done some observation to eliminate the possibility of CPU limiting... will get back on that.

Just quickly re-ran it partially. Total CPU usage was 9%, which is equivalent to just over one thread. Per thread activity on the system, I saw three threads on separate cores with anything above baseline activity, none of which were hitting max. So, this would not seem to be CPU limited in any obvious way.

 

14 minutes ago, Mark Kaine said:

Well you also should try a custom voltage / frequency curve because that reduces power draw significantly compared to just setting a power limit in my testing... 

 

I kinda think keeping the voltage low is key here, but seeing im almost at idle voltage already idk how much further it can go.

I agree that would reduce power, but I also consider that outside the scope of my test. There is some voltage/frequency curve already programmed into the card. It would be very time consuming to adjust that curve, and even then I consider it like an overclocked condition. What works for my specific sample might not work on any other. No one in mass production would tune it that close to the limit. We might see the shipped voltage as higher than needed at times, but that is to help cover for the case the product should be stable with any load you throw at it, and not just pick one or two to test with. If the GPU might be binned for better power efficiency, that would be a plus too, but again outside of what I can test at home.

[LogicalDrm]

-> Moved to Graphics Cards

[thepoorgamer]

Even if you coulden’t do 75 w, 120w would give you a lot of options such as using sata to pcie power adapters, like with the “this is a gaming pc?” video. That would be a more realistic optIon where you could extract a lot more performance while still being able to upgrade a prebuilt.