vivster said:
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.
|
GPU's for years have always executed two instructions per clock... Which is why we always used the:
Clock Rate * 2 instructions per clock * Shader Pipeline algorithm to determine theoretical floating point performance...
I.E. Xbox Series X.
Clock Rate = 1825Mhz * 2 instructions per clock * 3328 shader pipelines = 12,147,200Gflop or 12.14 Teraflops.
What nVidia has done here is made each pipeline branch out into two.
Shader counts and Flops have never been comparable, it was always a coincidence that they lined up in comparisons.
It's always been a hypothetical maximum rather than a real-world capability.
Case in point you could pick up a Geforce 1030 DDR4 variant and compare it to the GDDR5 variant, identical Gflops... But the DDR4 variant will be half the performance... Which just reinforces the idea that there is more to rendering a game than how many flops you have.
For example, AMD has always had more "flops" than nVidia... But in gaming that didn't amount to much, why was that? Because games do need things like Vertex and Geometry operations to also be performed, games may use integer texturing as well.
In short it meant that nVidia had better gaming performance... However.
If you were to throw a pure Asynchronous compute task at an AMD GPU, it will switch into another gear and it will beat nVidia hands down, which is why AMD GPU's were such lucrative buys during the crypto-craze, they were pure compute tasks and AMD's GPU's were built better for that.
vivster said:
I think that has less to do with the willingness of the console manufacturers and more with the ability of AMD. I'm sure they would've loved a feature similar to DLSS, but I doubt AMD has anything to offer on that part. Which kinda is a shame because DLSS would be absolutely perfect for consoles who already struggle with performance and image clarity. Of course they could've gone with Nvidia, but that would've been a real mess technologically, economically and politically. Less so for Xbox, but still.
|
Alternatives to DLSS exist.
Radeon Image Sharpening for example.
|
Then how would you explain the massive loss in efficiency of theoretical shader performance to real world from Turing to Ampere? Where's the bottleneck?
If I was Nvidia I would be incredibly insulted to have DLSS compared to RIS.
