By using this site, you agree to our Privacy Policy and our Terms of Use. Close
ElPresidente7 on 12 January 2017
Edit Thread
Soundwave said:
nomad said:

You 've said the GPU has more TDP headroom. No,  it does not.

That 1.5W is a misenterpretation of data. The test was conducted with an underclocked GPU. The test was not to show how much the GPU consumes but how effecient it can be. It shows that X1's GPU's performance at 1.5W is comparable to an iPad Air 2's(A8X) GPU which runs at 2.5W. It's unbelievable for a device that regurlarly hover at ~15W at load to only have a 1.5W GPU which takes up a significant portion of the SoCs size. The A57's are power hungry (which explains switch's heavy downclock) but not that hungry. The X1 GPU regulary perform over 2x that os the A8X in graphics benchmark to, assuming power effeciency linear (it isn't), that is 3W, but since it is nonlinear, lets just say 4W. Even then, I think that's being conservative, I wouldn't be suprise if it was >5W at full clockspeed.

 

Even at 5 watts, that's 100 GFLOPS/watt, if Nvidia was actually achieving that, everyone would be using Nvidia. 

You could have a 1.4 teraflop machine for just 14 watts, the XBox One die shrunk to 14nm/16nm (even smaller than the Tegra X1's 20nm) still consumes over 60 watts ... there's no way Nvidia has such magic up their sleeves, especially with a chip that is about two years old. 

No, this is incorrect. Power consumption increases quadratically with the clock frequency! A good GFLOP/watt ratio is very well possible with a very low clocked GPU or CPU.

Oh and btw. Mdave also showed that the CPU throttles the GPU and vice versa. His tests give a good reason also for the 1Ghz clock rate of the CPU. It seems that Nintendo just fixed the clocks of the CPU and GPU at a reasonable frequency that the system can handle maintain for a longer time without throttling.

If you read the thread, it seems we are still quite a bit away from having a whole picture of whats going on inside the switch. Nintendo typically does quite of customization to the chips. Even a 16nm process isnt out of the question. That would give room for quite a bit more souce.