Nvidia Ampere announced, 3090, 3080, 3070 coming later this year

[vivster]

eva01beserk said:
I guess im not the only one who saw a disparity with the TFlop number advertise and the performance gained advertised. did NVIDIA cards became less efficient with this gen? Does it have anything to do with TSMC abandoning them and the poor quality of samsung?

Does anybody know the size of these chips in comparison to touring? Just some rumors I heard that they are going with a huge chip to get more performance. Judging by the watt needed it seems that it might be the case.

I dont know guys, but the rumors seem sorta accurate and AMD seems to be going to give NVIDIA a run for its money. While I dont think they could possibly match the 3090, AMD did claim 2x of the 5700xt wich will be around of the 3080. If they price it right now that nvidea went first I see a big blow out in the future.

I've looked into it a bit. The big problem here is Nvidia using misleading numbers, which are technically true but have no bearing in real world applications.

Ampere is using a new kind of shader that is able to execute two instructions per clock. This is similar to hyperthreading in CPUs. However it's of course less efficient in real world applications than having 2 full shaders. Nvidia now proceeds to treat those new shaders as double the shaders from which they also derive the FLOPS. This brings us into a bit of a predicament because now the shader count and FLOPS are not comparable to Nvidia's own cards anymore.

So now you have 2 ways of looking at it. If you take the logical cores and theoretical FLOPS at face value you could say that the new shaders are less efficient compared to Turing, which would be true. But that kinda devalues the engineering that has actually been done. I would like to look at it differently and just half the proposed shader count and FLOPS and say that while they have not much increased from Turing, they have massively increased in efficiency.

As for the die size, we have hard numbers on that. To my personal surprise the GA102 used for the 3080 and 3090 is actually slightly smaller than the TU102 which was used for the 2080ti. At the same time it is almost twice as densely packed, which explains the massive performance increase.

TU102 - 754 mm², 24.7M/mm² (2080ti)

TU104 - 545 mm², 25.0M/mm² (2080)

GA102 - 627 mm², 44.7M/mm² (3090, 3080)

GA104 - 450 mm², 40.0M/mm² (3070)