12:45 karolherbst: ahh interesting... guess nvidia thought splitting up int and fp completly wasnt makeing all that much sense, so with ampere we have a fp32 and a fp32/int block :)
12:45 karolherbst: I guess that's some dual issuing magic again :)
12:45 karolherbst: why are they changing this stuff around so much? :D
12:49 mmenzyns: Maybe to have even better versability for all their future "technologies"? you tell me :D
15:41 HdkR: karolherbst: ampere is a very big GPU :D
15:41 karolherbst: HdkR: sure, I was just surprised that they brought back the mixed ALU cores
15:42 karolherbst: but it's fun that they get back to kepler level dual issuing, just that you now have a fp32 only core :p
15:44 HdkR: I'm surprised you get dual issue and a mixed pipeline
15:44 karolherbst: why would it matter?
15:44 karolherbst: dual issuing doesn't happen inside the cores, does it?
15:44 karolherbst: you can just schedule an instruction to two cores at once
15:44 karolherbst: uhh..
15:44 karolherbst: two instructions
15:46 HdkR: I would assume it happens in the frontend
15:46 HdkR: It makes sense to do both things at once. I just totally wasn't expecting it
15:46 karolherbst: well, they did it back with kepler already
15:47 karolherbst: but then they got rid of it again
15:47 HdkR: Which is why we get the silly 35TFLOP/s number with top end