Hahaha, you corrected my mistake and then made the same one yourselves. It would be correct to use N3 and N4. As for N4, it isn’t a custom process by Nvidia. Many others have long been making, or are still making, their chips on this node. However, N4P is a process purely tailored to Nvidia’s needs.

Roughly speaking, doubling the power gives around +25% performance on the same chip.

This is especially obvious on mobile chips, and it applies to others too. A 120W GPU is about 25% faster than a 60W one. This corresponds to the basis that power consumption grows exponentially, while frequency grows linearly.

N3 saves about 30% power at the same performance compared to N5. To reach similar performance 5070, you’d need roughly 175W TDP, whereas N1X has 140W TDP. To get 20% less performance, you’d need around 90W TDP, but here we have 140W TDP. Performance will be very close.

These two factors would work perfectly if this were just a mobile GPU, for example, the mobile 5090 on a 5080 chip with 2x lower TDP behaves similarly. But with an ARM SoC and limited memory bandwidth, there are additional factors that could negatively impact performance.

GDDR7 on the mobile 5050 compensates for the reduced TDP, making it even about 1% faster than the desktop 5050 on the same chip, but with GDDR6