Yeah, it doesn’t make a lot of sense to us for the Nintendo Switch to have such a GPU. 1 TFLOPS with a GPU clocked at 768 MHz? To get about 1 TFLOPS FP32 with 768 MHz there would need to be at least 640 CUDA cores. That would be a very large chip.

nintendo switch
via technobuffalo

Nintendo Switch can’t have a 1 TFLOP GPU

And 307.2 MHz in mobile mode for the Nintendo Switch? When’s the last time NVIDIA ran anything at such a low clock rate? More importantly, NVIDIA has very good power management that takes advantage of, among other things, controlling the clock. Why would they throw that out the window and lock developers to 307.2 MHz, as the article claims?

Why would they lock the CPU frequency at the docked frequency? Locking something lower might make some sense if they want to force a low power usage. But why force it to be run higher than it has to be? That just seems to prevent battery-saving in games that don’t need the full CPU speed when mobile.

If they have 1 TFLOPS FP32 then they have 1 TFLOPS FP32. It doesn’t matter how they get it. But this method doesn’t really seem to make sense. Even throwing out the clock locking stuff, we think 20nm would have to be a lot cheaper and have much higher yields to make it more economical to make a low-clocked SoC with 640 CUDA cores on 20nm over a higher clocked one with 320 or 384 CUDA cores on 16nm. Even if it made sense right now, would it still make sense a year from now when they will still be producing them. And probably making even more of them?

We think if they go with 20nm then they have lower performance. They aren’t going to put an insane number of CUDA cores on the chip to make up performance.