To follow up on this, I had a play around with MSI Afterburner, hoping to get a few data points, but basically I've only got one usable point. MSI Afterburner doesn't seem to allow you to set the GPU clock to anything below around 1.1GHz, which is a bit frustrating. Understandable, as it's an overclocking tool, not an underclocking tool, but still frustrating as it seems to be a bit of an arbitrary limit (the card itself has a frequency/voltage curve going down as far as 210MHz). It's also quite laborious to actually set a max clock, as you have to tweak a curve point-by-point to set about 50 individual values pixel-perfect.
Anyway, to test things, I did a short run (a couple of minutes) of Metro Exodus Enhanced Edition with ultra settings while logging stats via GPU-Z. I then just pulled the average GPU chip power consumption (ie excluding RAM or anything else) figure from the course of the run. My most reliable setting (reliable as in both clock and voltage were stable through the run) was 1155MHz, which ran at a voltage of 712mV, and the average chip power consumption over the run was 61.8W.
The RTX 3070 has 46 SMs, so that would put a very rough estimate of power consumption per SM at 1.343W at 1155MHz.
I feel I should list some caveats at this point:
- I don't know how accurate the GPU-Z "GPU chip power draw" numbers are
- We shouldn't expect power consumption to scale precisely linearly with SM count, there's other logic on there than just SMs
- This is the measurement of a single card, and my particular RTX 3070 may be particularly efficient, or particularly inefficient
- This is only one game, other games may consume more or less power
- My system of measurement isn't particularly scientific, I'm just looking to get rough numbers
That said, if we're looking at an 8nm chip, and all 12 SMs were running at 1155MHz, then the
very rough estimate of power draw would be 16.1W. This is just the GPU alone, so once you add CPU, RAM, etc, the full system power would be in the 20-25W range, which is quite a big increase over the original Switch. It's very much at the upper end of possibility for an 8nm chip, but it's actually a bit better than I'd expected. I would have said 1GHz was probably around as high as you could hit on 8nm. Again, this is only a very rough estimate, and real-world results on Drake may be very different.
It's difficult to say too much about what power consumption would look like at lower clocks without being able to test them directly. There's a temptation to think that there would be a big increase in efficiency by going lower, but there is reason to doubt that. In particular, the idle voltage of this card is 681mV, where it will drop down to 210MHz. That's not a huge drop from the 712mV we're seeing at 1155MHz, particularly when you consider the card will go up to around 1200mV at peak clocks. This would suggest that, while there's definitely some power to be saved by clocking lower, there isn't room for big voltage reductions and the associated boost to efficiency.