Crazy idea: power testing (40 series) for custom SFF cooler

[Hewligan]

I would really appreciate if anyone who owns a 4070 Ti, 4080, or 4090 could help with the following test (below). This discussion started on the NFC discord, and the intention is to see how much power we could pack into a tiny case (the Skyreach 4 Mini, specifically, at 5.5L). I own 2 of these cases, and the last GPU I had installed was the EVGA XC 3060 Ti, which just fits given the 2 slot and 215mm constraints.

With the 40s series being a bit underwhelming from the pricing perspective, and the huge coolers that seem to come with even modest cards, I just ignored it. But then someone started pushing the idea of a custom cooler on a heavily power limited and undervolted RTX 4090 fitting in the above case.

So here is the added spiciness for this: powering from 2 x HDPlex 250 watt GaN units (they can be daisy-chained, with one powering the GPU, and the other powering the motherboard, CPU, etc.). So a total headroom of 250w + whatever the PCIE slot provides. The HDPlex can handle some spikes before tripping, but ideally you wouldn't want to regularly throw more than 250w average, and 300w with spikes at it to keep it stable.

So the request is: if you own a 40 series card, could you please use MSI Afterburner to provide 2 sets of results (FPS and GPU Power Usage) using Kombustor with the "BENCHMARK: preset 1440"?

1) At stock
2) Undervolted and power limited to get a power draw as close to 250w for the GPU as possible (if you could post your voltage curve that would be great).

The intention is to work out if the 4070 Ti, or the 4080, or the 4090 are materially different in FPS at such a constrained power point - before we investigate spending lots of money and time making a custom cooler to handle the 250w of heat output (which in theory the case can handle - it was used with the mini 1080 Ti by many people in the past).

I appreciate this is niche, and also time consuming, but if you could help it would be greatly appreciated.

Many thanks,

Hewligan

 

[pastymuncher]

You are going to need a hell of a lot more than 250w for a 4090 even when undervolted. My 3070 pulls more than that when underclocked/undervolted and that's with just over 100w saved from undervolting. Modern cards have ridiculous power requirements. Even if by some miracle you could get near 250w you would have underclocked and undervolted it so much it would make it pointless to have such a card and would be better off waiting for the lesser cards to come along.

 

[Hewligan]

You can undervolt and then power limit the card, and still achieve high performance compared to prior gen.

Derbauer demonstrated that in a video already.

 

[Wing-Man]

Derbauer demonstrated that in a video already.

show use where Derbauer got a 4090 to pull only 250w underload

 

[Hewligan]

The RTX 4090 Power Target makes No Sense - But the Performance is Mind-Blowing

Support me on Patreon: https://www.patreon.com/der8auer---------------------------------------------------------Save 10% on your iFixit purchase: DER8AUER10h...

youtu.be

16:10 for the chart and discussion

 

[stellasuppa]

If you power limit a 4090 beyond 60% the performance drops off quite a bit. The 4080 and 4070ti would certainly be able to perform under them power target constraints.

 

[Alt0153]

You are going to need a hell of a lot more than 250w for a 4090 even when undervolted. My 3070 pulls more than that when underclocked/undervolted and that's with just over 100w saved from undervolting. Modern cards have ridiculous power requirements. Even if by some miracle you could get near 250w you would have underclocked and undervolted it so much it would make it pointless to have such a card and would be better off waiting for the lesser cards to come along.

using a 4090 at 1440p generally keeps it below that

 

[momo56]

Had a quick go with a 4090.

136 FPS ~ 248W
176 FPS ~ 436W

Spoiler

 

[Wing-Man]

Had a quick go with a 4090.

136 FPS ~ 248W
176 FPS ~ 436W

Spoiler

this is very intresting can you do the same two tests at 4k

 

[Hewligan]

Had a quick go with a 4090.

136 FPS ~ 248W
176 FPS ~ 436W

Wow - thank you so much for actually doing the test. I had half given up hope of getting a response.

 

[CubanLegend]

I'm the crazy one proposing this idea, nice to make yall's acquaintance.

Had a quick go with a 4090.

136 FPS ~ 248W
176 FPS ~ 436W

Spoiler

thank you for this test! this looks promising.

If you power limit a 4090 beyond 60% the performance drops off quite a bit. The 4080 and 4070ti would certainly be able to perform under them power target constraints.

Performance doesn't seem to drop that dramatically as per the Der8auer tests, and other Youtuber's who ran similar tests.

The idea is that if I limit the 4090 to 60%, if in Timespy I get 284w, and in KOMBUSTOR at 44% power limit, i get under 250w.. that logically in real-world gaming scenarios, it should be even less power being used at those power limits.

 

[tommybhoy]

If a 70ti cost £800 and the 90 costs £1600, how many hrs of gaming is required to break even?

 

[Hewligan]

If a 70ti cost £800 and the 90 costs £1600, how many hrs of gaming is required to break even?

Depends what value you ascribe to your hobby, and how much cash you have to allocate to your hobby.

 

[bimbleuk]

If I remember I might be able to post something later on a 4080 FE.

I have my 4080 UV anyway and it seems to hover around low 200W most of the time playing games at 3440 x 1440. I'm sure when I enabled DLSS quality while playing Hogwarts with everything Ultra + RT it dropped below 200W but I'd have to double check that. Wasn't above 50C anyway this morning when I had a quick go before work.

 

[Hewligan]

If I remember I might be able to post something later on a 4080 FE.

Thank you. That would be fantastic.

 

[tommybhoy]

Depends what value you ascribe to your hobby, and how much cash you have to allocate to your hobby.

Although an undervolted 90 might still be slightly more efficient than an undervolted 70, but at what point do you stop?

I realise that it's a hobby and you want to achieve a goal, it wasn't a dig, more wondering what the figure to my question is.

 

[Hewligan]

I guess it depends on the outcome of the tests. My expectation is that the 4070 Ti will come out very close at 1440p to the 4090 when power constrained to this level, but CubanLegend disagreed with me, hence the request here before any actual money is spent. At the moment we are trying to work out where that inflection point that offers the best balance between power efficiency and cost sits. I think CubanLegend leans more towards maximising performance per litre. I am more on the efficiency side, but without some numbers we are just guessing at the moment.

 

[bimbleuk]

I remembered correctly running Hogwarts Ultra 3440 x 1440 + RT with no DLSS used ~220W. Enabling DLSS quality dropped power to ~180W with >60 fps maintains decent qaulity.

Here are some results from running Kombustor using the 1440 pre-set.

Seems between 70-80% power limit is a good place to aim for. Clock speed was mostly stable until 70% where you can see a ~200 MHz drop from around 2775MHz on my 4080 FE.

Example curve. Though my "curve" just lifts the voltage limit to my target clock speed and then flatten across. I used .975 mV & .950 mV and then limited power as well.

Spoiler: Example voltage curve

 

[Hewligan]

I remembered correctly running Hogwarts Ultra 3440 x 1440 + RT with no DLSS used ~220W. Enabling DLSS quality dropped power to ~180W with >60 fps maintains decent qaulity.
[…]

Thank you!!! Wow, so close to the 4090 at those power levels. Sounds like it starts to suffer sub 250w, but there is a sweet spot that is just a few % from the 4090.

 

[CubanLegend]

If a 70ti cost £800 and the 90 costs £1600, how many hrs of gaming is required to break even?

Although an undervolted 90 might still be slightly more efficient than an undervolted 70, but at what point do you stop?

I realise that it's a hobby and you want to achieve a goal, it wasn't a dig, more wondering what the figure to my question is.

Honestly its not really about breaking even, as Hewligan said.. "its about attaining the highest possible FPS per liter" while not thermal throttling (staying below the TJmax of the GPU), fan speed and fan sound isn't a worry here.

I remembered correctly running Hogwarts Ultra 3440 x 1440 + RT with no DLSS used ~220W. Enabling DLSS quality dropped power to ~180W with >60 fps maintains decent qaulity.

Here are some results from running Kombustor using the 1440 pre-set.

Seems between 70-80% power limit is a good place to aim for. Clock speed was mostly stable until 70% where you can see a ~200 MHz drop from around 2775MHz on my 4080 FE.

Example curve. Though my "curve" just lifts the voltage limit to my target clock speed and then flatten across. I used .975 mV & .950 mV and then limited power as well.

Spoiler: Example voltage curve

Thank you for submitting your results. It's so intersting how when you induce an Undervolt your performance per power limit seems to increase ever-so-slightly.

So it seems a 4090 FE at 44% power limit (250w) is within a FEW percent of a 4080 FE between 100%-80% power limit (w/a 950mV undervolt).

honestly I'm even more curious of if we can remove the 250w limit we've imposed on this crazy idea. If we used a different AC-DC/DC-DC setup, we may be able to handle more than 250w and even have room for transient spikes well above that, which are to be expected and need to be addressed.

Thank you all for your results so far!