sc94597 said:
The reason why you can't compare the power-efficiency of a (down-clocked) RTX 2050 to the power-efficiency of a T239 is because they are different chips that target different power profiles, have different SOC (not micro, SOC) architectures, and therefore different power-efficiency curves. The comparison makes sense from a performance perspective (even if rough) but not the power-efficiency perspective. There also is the matter of the different CPUs. This is true even if the GPUs are both Ampere. I'll provide an example of why this is the case, Consider the RTX A2000 and RTX 3050. They both have a GA106 die. The A2000, however; has 3328 cores clocked at 562* - 1200 Mhz while the RTX 3050 has 2560 cores clocked at 1552 - 1777 Mhz. The A2000 slightly performs better than the RTX 3050 in games and compute. Yet the RTX A2000 pulls about 70W at full utilization and the RTX 3050 pulls about 115W - 130W at full utilization in its original form. Now this is for cards that have the same GA106 (just cut and clocked differently.) *Notice the 562Mhz here, this likely isn't a coincidence, that probably corresponds with an actual power-state. As to the exercise of halving the TFLOPs, you'll notice that the 1.84 Tflops PS4-equivalent performs worse than the 1.39 TFlops GTX 750ti on that benchmark. TFlops and real-world performance are only loosely related, and you really can't extrapolate one from the other when comparing the performance across architectures. Also the curves are not linear, so halving the GPU's power utilization doesn't necessarily mean halving its performance. The curves do become near-linear after some low-point, but there is reason to believe that a T239 has that point shifted leftward compared to an RTX 2050, due to their different power-targets. As an example, this is what a power-performance curve of an A5000 (light-blue) looks like. or the power-performance curve of a Steam Deck Oled.
Notice that at 6.5W it is retaining 83% of the performance as it has at 12W (1500Mhz vs. 1800Mhz, all else equal.) Power utilization nearly halved (6.5/12 ~ 55%), but theoretical performance** didn't (1500/1800 ~ 83%.) **TFlops are proportional to clock-rate and core-count. Core-count is a constant here. |
All chipsets designed for portable hardware like handheld consoles and laptops target efficient power profiles, there is nothing new in the Switch 2 in that regard. You have a Samsung 8Nm fabrication process which is more like 10Nm in reality and you have only 20Wh to use. Some of the newer PC based handheld consoles using Z2 extreme also have power efficient profiles but they use a 4Nm fabrication process and have up to about 99Wh for battery on the largest battery models. A cheap fabrication process and a low capacity battery do not allow for miracle performance levels. Surely your argument if anything shows Nintendo are best advised to clock low on the Switch 2 to get maximum battery life so if anything reinforces my believe about the real performance level of the Switch 2. The efficiency curve for performance per watt is set in the design, it would have been designed from the ground up to be efficient at its stock frequencies and more power would only be wasted and achieve poor performance levels for the increased frequencies/wattage so for these types of designs reducing power will have a noticeable drop in performance just like if you set your laptop to a 83% performance level you don't suddenly get twice as long out of the battery that isn't how it works. These are already power optimised designs. Yes you can throw more power at them and achieve some small performance gains but they are not designed this way. The Switch 2 isn't capable of delivering superior power efficiency with a dated fabrication process and small capacity battery.
The power efficiency of chips is completely linked to the fabrication process, yes other factors might have a small element in power efficiency but ultimately the fabrication process dictates power efficiency. Of course in the final designs other parts of the design are a huge factor too like the screen, amount of memory, wifi chip etc but you can't get away from power efficiency being linked to the fabrication process. So the idea that a dated effectively 10Nm fabrication process can compete with a 4Nm fabrication process is not realistic for power efficiency.
I'm personally quite confident that the Switch 2 will be achieving a sub 1 Teraflop performance level in portable gaming to achieve around 2 hours but the final numbers will probably be revealed in a month or so. Ultimately if Cyberpunk has a native resolution as low as 360p this surely is an indicator of GPU power level available. Lets not forget the Deep Learning Super Sampling is there to upscale amazingly and the games will already be optimised so they work exceptionally well with that upscaling technology. Hence less artifacts than doing the same on PC. It's a very clever design that doesn't need to be that powerful. Yes if Nintendo had coughed up for a better fabrication process which could have had much longer battery runtime for the same performance level but they didn't. It is what it is.
Lets not forget either that with this cheaper fabrication comes more heat and therefore cooling requirements which is another reason to clock low which they appear to have done according to that video. Nintendo designs very reliable hardware generally and doesn't want to cough up for warranty claims if they can. They are completely focused on margins and profit.
