RTX 4070 vs RTX 4070 Ti power scaling

[10tothemin9volts]

RTX 4070 vs RTX 4070 Ti power scaling
You may see also my previous power scaling post.

• After -35% power, the 4070 non-Ti drops much quicker. What are the reasons, maybe a better binned chip?
• After -5% power, the 4070 Ti scored slightly higher or not worse.

• Even after -50% power (285W*0.5=142.5W), the 4070 Ti scored higher than the 4070 non-Ti at full its 200W TDP.

Couldn't wait for RTX 5000 series (rumor: RTX 5090 release in December and more affordable versions possibly only many months later), so upgraded from RTX 4070 to a used 4070 Ti. I didn't went for a 4070 Ti Super with its 16GB VRAM since I couldn't find a used one for much cheaper (it came out indeed: "too little, too late"). Also when I bought the 4070, it was the first NVIDIA card I used on Arch Linux and I wasn't sure if it was worth it spending much on a NV card, now I know that there are no issue. The perf increase will be useful for the 1-50% FPS lows in CS2:ZE since with 64 players the average FPS on some maps is like only ~30 FPS). And of course: No HDMI 2.1 on Linux on AMD is a deal breaker.

 

[Macro Device]

After -35% power, the 4070 non-Ti drops much quicker. What are the reasons, maybe a better binned chip?

The 4070 has 77% SMs and 100% VRAM at 100% speed at 70% power. That's why it starts getting stunted "earlier" as it has less power per SM from the very beginning.

3090 has 24 VRAM chips at ~140 W power draw at full throttle. Yes, 140 W only for memory.
4070 and 4070 Ti have 4 times less VRAM modules. I don't think they're much more efficient so let's say they consume ~30 W.

Now we are left with ~170 W for a GPU in 4070 and ~255 W for a GPU in 4070 Ti. 170/46 = 3.7 W per SM. 255/60 = 4.25 W per SM.

To "equalise" them, the 4070 Ti should run at 220 W die-only, resounding 250 W total power. Which is around 88% TDP.

Better binning also might've been the case but I doubt it matters this much all things considered.

 

[RTX4070TI]

70ti has the best ad 104 chips thats why.

 

[Taraquin]

If you undervolt a 4070 you find the reason quickly. At 0.91v you get about 65-70% of powerconsumption at 5-10% lower powerdraw. If you drop below 0.91v vram downclocks from 21 to 10GHz. This causes a massive performancedrop. I havent trsted 4070ti, but maybe it has a different voltagecurve/vram downclockpoint?

 

[10tothemin9volts]

Thanks for the infos.

[..]

Very useful and interesting infos, which make sense, and may confirm why the 4070 Ti's heat output feels too much (it shouldn't be the GPU's VRM as it's not the Extreme model, which has like a 40 Watt idle power consumption, and it was more like ~20 Watt). The heat output is the first thing I negatively noticed. Also, when touching the GPU's cooling fins, they often felt very hot, almost too hot to the tough. Maybe one can use your infos to choose a less heat-outputting GPU next time. I sold my 4070 Ti due to the heat it was outputting into my room, it was kinda crazy tbh, despite 285W vs 200W doesn't seem that much of a difference (basically the 4070 non-Ti feels very cool compared to the Ti).

24 VRAM chips at 140W is maybe why the RTX 4000 series has 2GB VRAM modules. If the VRAM chips use so much power, it's time for 3-4GB per chip (And preferably no increase in power consumption per chip. Wonder how 1GB vs 2GB VRAM modules compare in power consumption.) (unfortunately GDDR7 only has a bandwidth increase), now that running local LLMs (think ChatGPT, but locally on your PC) is a thing and actual demand for more VRAM is there.

PS: The Superposition settings used: Custom, 2560 x 1440 (QHD), Extreme, High and the raw data is:

percent,4070,4070ti
100,5343,7240
95,5308,7267
90,5265,7249
85,5213,7245
80,5120,7176
75,5076,7030
70,4977,6912
65,4865,6746
60,4514,6568
55,4100,6317
50,3514,5979

Need a better benchmark (or use a frame recording software) which also outputs the 1% and 0.1% percentiles to see if reducing the power effects these.

 

[tabascosauz]

depite 285W vs 200W doesn't seem that much of a difference (basically the 4070 non-Ti feels very cool compared to the Ti).

So a 40%+ increase in power consumption isn't "that much" to you?

From my limited adventures in undervolting with 4070 Ti, it also doesn't seem to achieve as good results as 4070 did. Both of which also seem to be less impressive than the new 4070 Super at undervolting (at a glance).

I have to assume that binning is just that much better on the new GPU, for there to be that much of a difference in efficiency. The 3070 Ti was much the same story in comparison to the 3070, although there most of the difference could be blamed on first gen GDDR6X.

 

[10tothemin9volts]

So a 40%+ increase in power consumption isn't "that much" to you?

Yes exactly, only +42.5 % doesn't justify why I also added: "(basically the 4070 non-Ti feels very cool compared to the Ti)". The now sold 4070 Ti felt like, idk, a 400W-600W TDP card or something (never had one, just guessing). Maybe the VRM was broken, but probably not. Even when limited to 150W it output lots of heat compared to a 150W 4070 non-Ti. But maybe the 42% was just enough to feel so my skin like 300% (skin temperature sensitivity doesn't scale linearly). Guess these are the things that are not conveyed by reviews. At least my 4070 Ti blasted, what I would describe as a heat wave.

 

[tabascosauz]

Yes exactly, only +42.5 % doesn't justify why I also added: "(basically the 4070 non-Ti feels very cool compared to the Ti)". The now sold 4070 Ti felt like, idk, a 400W-600W TDP card or something (never had one, just guessing). Maybe the VRM was broken, but probably not. Even when limited to 150W it output lots of heat compared to a 150W 4070 non-Ti. But maybe the 42% was just enough to feel so my skin like 300% (skin temperature sensitivity doesn't scale linearly). Guess these are the things that are not conveyed by reviews. At least my 4070 Ti blasted, what I would describe as a heat wave.

Hmmmm interesting. I've not noticed anything peculiar between the 3070 Ti (TUF), 7900XT (MBA), and 4070 Ti (Eagle), all at similar power draw. Not that I would really desire putting my hand on any of them at full 3D load, they get rather warm. You sure nothing was amiss on that used 4070 Ti?

 

[10tothemin9volts]

RTX 4070 vs RTX 4070 Ti power scaling and power efficiency​

I added the power efficiency graphs, which make it easier to read (can't edit my first post):

[...] Not that I would really desire putting my hand on any of them at full 3D load, they get rather warm. You sure nothing was amiss on that used 4070 Ti?

When the GPU was idle, I think the fins were more often hot to the touch, sometimes not. When the GPU was utilized I think it was not hot to the touch. Thinking about it, I'm not sure about it anymore (GIGABYTE GAMING OC v1), maybe something was indeed wrong with that GPU, I bough it used after all. Or maybe the VRM drew just too much and that model was replaced with a v2 (but the v2 is also more compact, iirc).

 

[..0]

i run my 4090 at 50% power limit, also added 150mhz. seems like a quick and reasonable tuning for that gpu.
i also set my powerlimit to 133% to check and most of the time i get not extra performance. my 12700k is a bottleneck in most games i play.

 

[#22]

Typical view for Nvidia GPUs - cut off 1/3 of power to loose 10% performance. Whopping reduction of heat and noise for unnoticeable fps difference. Sad reality of modern GPU scene and answer to what happened to overclocking headrooms we used to see :F With CPUs it's even more absurd. I wonder if it's the same with Radeons.

 

[tabascosauz]

RTX 4070 vs RTX 4070 Ti power scaling and power efficiency​

I added the power efficiency graphs, which make it easier to read (can't edit my first post):

View attachment 346803

When the GPU was idle, I think the fins were more often hot to the touch, sometimes not. When the GPU was utilized I think it was not hot to the touch. Thinking about it, I'm not sure about it anymore (GIGABYTE GAMING OC v1), maybe something was indeed wrong with that GPU, I bough it used after all. Or maybe the VRM drew just too much and that model was replaced with a v2 (but the v2 is also more compact, iirc).

That doesn't sound right. I have the v1 Eagle with the 301mm cooler (as opposed to v2 Eagle with the 261mm cooler) and nothing is scalding hot at any point in time. We should have a similar if not identical PCB. Doesn't really make sense for the VRM to be the culprit, but clearly the power draw had to be there for the heatsink to get that hot.

Although, GPU does remain fanless even as high as 80-90W, at which point the finstack won't be cold anymore.

 

[10tothemin9volts]

[...] I wonder if it's the same with Radeons.

I did a comparison on the Radeons. RDNA3 doesn't power scale that well compared to the RTX 4000 series (at least according to the data -- not my data). The chiplets could be a reason for it, as moving data supposedly consumes by far the most power. The lithography (TSMC N5 (GCD)+N6 (MCD) vs NV using TSMC 4N) may also be a reason. Comparing RDNA2 vs RDNA3 non-chiplet/monolithic RX 7600 (XT) based variants could tell more.

That doesn't sound right. I have the v1 Eagle with the 301mm cooler (as opposed to v2 Eagle with the 261mm cooler) and nothing is scalding hot at any point in time. We should have a similar if not identical PCB. Doesn't really make sense for the VRM to be the culprit, but clearly the power draw had to be there for the heatsink to get that hot.

Although, GPU does remain fanless even as high as 80-90W, at which point the finstack won't be cold anymore.

That's good to know. In idle my was mybe fanless for 30s. I guess then something is with that GPU, someone must have sold it, as used, for a reason.

 

[#22]

I did a comparison on the Radeons. RDNA3 doesn't power scale that well compared to the RTX 4000 series (at least according to the data -- not my data). The chiplets could be a reason for it, as moving data supposedly consumes by far the most power. The lithography (TSMC N5 (GCD)+N6 (MCD) vs NV using TSMC 4N) may also be a reason. Comparing RDNA2 vs RDNA3 non-chiplet/monolithic RX 7600 (XT) based variants could tell more.

Thanks for the link! I find power scaling analysis interesting.

 

[tabascosauz]

I did a comparison on the Radeons. RDNA3 doesn't power scale that well compared to the RTX 4000 series (at least according to the data -- not my data). The chiplets could be a reason for it, as moving data supposedly consumes by far the most power. The lithography (TSMC N5 (GCD)+N6 (MCD) vs NV using TSMC 4N) may also be a reason. Comparing RDNA2 vs RDNA3 non-chiplet/monolithic RX 7600 (XT) based variants could tell more.

That's good to know. In idle my was mybe fanless for 30s. I guess then something is with that GPU, someone must have sold it, as used, for a reason.

I see what you mean now, that doesn't sound normal at all. Mine remains fanless at such high wattage that to get the behaviour you're seeing purely due to ambient temp alone, I'd have to live in a volcano. It's permanently fanless at idle - even when at forced max VRAM clock (4K120 + 1440p165 + 1440p165), I still can't get fans to spin, it's only about 30-40W.

And even at the very threshold of fans spinning up (80-110W), the heatsink is still only warm, not dangerously hot to the touch.

Do you have any means of measure entire system power draw? ie. an AX1600i? UPS? kill-a-watt?