I think they also tend to forget the increase in transistor density. It uses more power, because there is more to power. And those increased memory speeds don't come free either.
If you want low TDP/max efficiency, then you look at the binned mobile chips. The desktop GPUs are really the leaky cast offs.
All GA104 6144 Cuda Cores:
RTX 3070Ti = 290W @ 21.75 TFLOPS
RTX 3080M = 150W @ 21.01 TFLOPS
RTX 3080M = 80W @ 16.59 TFLOPS
So you can cut the power in half and get the same rough performance, just need the right chip.
Those RTX3060 coming out of China that are actually re-tooled laptop GPUs are actually really interesting, it would be cool to have some super efficient options like that. Something I might actually pay extra for.
That or I want motherboards to start accepting MXM cards, or whatever future equivalent. I still think it would be really cool to have a GPU mount on like the back of a motherboard or something. No need for riser cables in SFF chassis, just slap a heatsink on both sides of the motherboard, or sandwich between two water blocks.
If you want low TDP/max efficiency, then you look at the binned mobile chips. The desktop GPUs are really the leaky cast offs.
All GA104 6144 Cuda Cores:
RTX 3070Ti = 290W @ 21.75 TFLOPS
RTX 3080M = 150W @ 21.01 TFLOPS
RTX 3080M = 80W @ 16.59 TFLOPS
So you can cut the power in half and get the same rough performance, just need the right chip.
Those RTX3060 coming out of China that are actually re-tooled laptop GPUs are actually really interesting, it would be cool to have some super efficient options like that. Something I might actually pay extra for.
That or I want motherboards to start accepting MXM cards, or whatever future equivalent. I still think it would be really cool to have a GPU mount on like the back of a motherboard or something. No need for riser cables in SFF chassis, just slap a heatsink on both sides of the motherboard, or sandwich between two water blocks.