After spending hours looking at as many mobile devices with large GPU's as I could, especially anything in the realm of 17 billion transistors, one thing became very clear. Yes, it is possible today. You can make a small circuit board with a large chip. The biggest design challenge is the cooling solution for such a large chip, because the products out there have high clocks and high TDP.
This took all my dream clocks for T239 and crushed them.
I think based on what we are seeing from the other Orin products in terms of starting clocks, combined with the realities of putting anything similar into a mobile form factor, we should prepare ourselves for very very low clock speeds.
Nintendo will consider:
-keeping costs down and avoid any overly expensive cooling solution
-will prioritize battery life over performance
-will be conservative with thermals, especially in handheld mode, as they can't risk burning kids hands
The notebook line of GeForce 30 series chips are clocked from 713-1740 MHz (including boost modes) and consume from 35-150W.
I now believe Max docked clocks to be under 939 MHz, possibly as low as 765 MHz.
Handheld clocks under 384 MHz. If they are able to figure out a form of back-compat and run Maxwell GPU code on Ampere, then a mode that disables 83% of the GPU and run the rest at 384 MHz makes sense.
That works out to under 2.9 TFLOPS docked and given I believe half of the GPU will be normally be off while in handheld, under 0.6 TFOPS undocked.
The math:
Docked 2 * 1536 * 939 / 1,000,000 = 2.884608 TFLOPS
Undocked 2 * 768 * 384 / 1,000,000 = 0.589824 TFLOPS
This took all my dream clocks for T239 and crushed them.
I think based on what we are seeing from the other Orin products in terms of starting clocks, combined with the realities of putting anything similar into a mobile form factor, we should prepare ourselves for very very low clock speeds.
Nintendo will consider:
-keeping costs down and avoid any overly expensive cooling solution
-will prioritize battery life over performance
-will be conservative with thermals, especially in handheld mode, as they can't risk burning kids hands
The notebook line of GeForce 30 series chips are clocked from 713-1740 MHz (including boost modes) and consume from 35-150W.
I now believe Max docked clocks to be under 939 MHz, possibly as low as 765 MHz.
Handheld clocks under 384 MHz. If they are able to figure out a form of back-compat and run Maxwell GPU code on Ampere, then a mode that disables 83% of the GPU and run the rest at 384 MHz makes sense.
That works out to under 2.9 TFLOPS docked and given I believe half of the GPU will be normally be off while in handheld, under 0.6 TFOPS undocked.
The math:
Docked 2 * 1536 * 939 / 1,000,000 = 2.884608 TFLOPS
Undocked 2 * 768 * 384 / 1,000,000 = 0.589824 TFLOPS