Zach_01Not every cuda core is created equal. Its like trying to compare different cores from different CPUs.

-------------------------------------------------

Its easy if you know the specifics. According to Jensen a process unit to be called cuda core must be able to execute FP32 instructions.

Pre-Turing architectures every cuda core was capable of 1 Int or 1 FP instruction, and cannot execute both.

This has changed in Turing. In Turing every cuda core was performing exclusively FP instructions and the number was 4352. For Int instructions Turing had another 4352 units that was not counted for cuda cores (see Jensen)

Ampere 3080 now has (4352x2) 8704 cuda cores and the 4352 of them are executing exclusively FP instruction and the other 4352 are executing Int or FP instructions, but not both.

See?

AusWolfSo basically Int cores count as a cuda core now, while in Turing, they were just an extra unnamed component next to the FP cores? Interesting.

Zach_01Process units are called cuda cores, always was, as long as they can execute FP instructions, despite the Integer capability.

In Ampere all units can process FP but half of them can process Int. Turing had cuda cores only capable of FP, and other units (same amount) only capable of Int that are called... Int processing units, but not cuda cores.
Its the FP capability (do or dont) that defines a unit to be (called) cuda core or not.
Its Jensen on the work!

AusWolfAh, I see. The cynic in me starts to think that the term 'cuda core' is just another marketing BS (8000 is better than 4000, right?) However, it doesn't change the fact that Ampere offers minimal (if any) efficiency gains over Turing, which is sad.