Can anyone distill down into simple terms what we should realistically be expecting performance-wise from this machine based on what we know so far?
Notwithstanding uncertainty about RAM, and I will assume the chips will run clocked to their standard clock which is perhaps possible when it is docked, the power question boils down to three uncertainties:
1. What Tegra chip will be used (Tegra X1 or Tegra Parker)?
2. How does the power measurement method (known as FLOPS) translate between AMD and NVIDIA cards (NVIDIA's FLOPS tend to be 'stronger', i.e. the communication allows for the FLOPS to be realised more efficiently, giving, and I found this ratio in another Neogaf thread, a ratio of approximately 4:3 in favour of NVIDIA, and this translates approximately one on one to an increase in FLOPS as compared to AMD. We want the AMD FLOPS, because we talk about FLOPS for Xbox One and PS4 as well, being 1.31(?) TFLOPS and 1.84 TFLOPS, but they use AMD rather than NVIDIA).
3. There are two ways (actually, there are more, but without loss of much accuracy we can say there are two) to do computations: FP32 and FP16. The first is slower but more accurate and the second is faster (roughly twice, in fact) but less accurate. Finding a balance between accuracy and speed in this method can increase the FLOPS rate (using only FP16 would give twice the number as compared to using only FP32, for example).
Tegra X1 has 512 GFLOPS for FP32, and Tegra Parker has roughly 750 GFLOPS FP32. If we could, for example, use FP16 for 1/3 of all computations, then we gain a 20% increase in FLOPS.* So you see there can be a significant increase. Xbox One and PS4 cannot use this so-called mixed-precision computation (PS4 Pro
can, that's where the rumours about PS4 Pro doing 8.4 TFLOPS come from), so the Switch has a potential advantage in power in this regard.
Let's do a calculation: assuming the Tegra X1, we have 512 GFLOPS of NVIDIA FP32. Assuming (with no particular reason for assuming this, but some more technically-schooled GAFfers called it plausible, but it differs on a game-by-game basis) that 1/3 of the computations can be done in FP16 (which, as I mentioned, results in a 20% gain), we can compute the comparison between the two as follows:
Switch power = 512 * 1.20 * 4/3 = 819 GFLOPS. (the 4/3 is that NVIDIA to AMD ratio I mentioned before) This resultant number is how the Switch actually compares in power to PS4 and Xbox One (which are, respectively, 1.84 TFLOPS and 1.31(?) TFLOPS, remember 1 TFLOPS = 1000 GFLOPS).
If, on the other hand, the Switch uses a Tegra Parker, then the power will be (with the usual caveats that we are guessing a lot of numbers):
Switch power = 750 * 1.2 * 4/3 = 1200 GFLOPS = 1.20 TFLOPS. So, you see that using the latter setup, the Switch could be very close in power to the Xbox One. Remember, though, that the gain from FP16 computations is just a guess, as well as the ratio between NVIDIA and AMD FLOPS (the effect, though, is very real, just not numerically determined).
About the clock I mentioned: there is a standard clock value (Tegra X1 has it at 1 GHz), and the FLOPS rate scales linearly with this clock value (so, halving the clock value will half the FLOPS rate). In handheld mode, the clock value will go down as this saves heat production and ergo battery life. In dock mode, however, active cooling could possibly allow the chip to run at full clock speed and therefore allow the power values I determined above.
Disclaimer: The info I produce here is produced by someone who is not a computer engineer (yet), so there might be something wrong in my explanation. If someone spots an error in my explanation (remember, though, that this is purely a FLOPS determination: we simply don't know how RAM and other things will play into the equation), please correct me.
TL;DR/Conclusion: Depending on many factors, the Switch can possibly be very close to the Xbox One for at least a number of games, but that does assume lot of things we simply do not know, and things that often depend on a game-by-game basis. On the other end of the spectrum, though, the power could possibly be roughly half of the Xbox One, so you see there is a lot we do not know and a large margin for errors.
*: See Thraktor's post (#1551) to see how this gain can be calculated.
Edit: For those interested, I converted Thraktor's system of equations into a nice and simple calculation. Take a specific ratio (I will showcase the ratio FP32 : FP16 = 2 : 3). Do the following:
total = 2 * FP32 + FP16 = 2*2 + 3 = 7.
Now divide FP16 by the total:
gain = FP16 / total = 3/7 ~ 0.43. You can check by solving Thraktor's system that this result pans out every time. Here is the proof:
Assume FP32 : FP16 = n : m.
We must prove that FP32 + FP16 = FLOPS * (1 + m/(2n + m) ) (i.e. the new power is the original plus the fractional gain m/(2n +m) which is the fraction I described above).
Thraktor's system says:
FP32 + (FP16 / 2) = FLOPS
FP32 / FP16 = n/m. => FP16 = FP32 * m/n
FP32 + (FP32 * m/n) / 2 = FLOPS
FP32 (1 + m/(2n) ) = FLOPS
FP32 * ( (2n + m) / (2n) ) = FLOPS
FP32 =FLOPS * (2n) / (2n + m).
FP16 + FP32 = FLOPS * ((2n) / (2n + m)) * (1 + m/n). = FLOPS * ((2n) / (2n + m)) * ( (m + n) / n) = FLOPS * ( (2n * (m + n) / (n * (2n + m)) = FLOPS * (2*(m + n) / (2n + m)) = FLOPS * ( (2n + m) / (2n + m) + m / (2n + m) ) = FLOPS * ( 1 + m / (2n + m) ).
And that is what I had to prove (QED, as they say).