OTOY Forums

2slot nvlink (for quadro turing)

Stupid me (o;

Of course the RTX3080 can't be twice as fast as the RTX2080TI (o;

RTX3080 has twice the CUDA core amount, but the RT core amount is the same...

Therefore I can assume that doubling CUDA cores takes 60% of the speed improvement, whereas RT cores 40%.


So a RTX3090 with 10496 cores has 2.412 times more CUDA cores as RTX2080TI, and 60% of that gives a point increase of around 300 ( (10496 / 4352 * 355 - 355) * 0.6)...
And the RT core increase from 68 to 82 gives a point increase of around 29...( (82 / 68 * 355 - 355) * 0.4)

So my assumption is that the RTX3090 should be 684, or 1.93 times faster than a RTX2080TI...
but with the base clock being a little higher, it should be at least factor 2...

In another thread here I read that OctaneBench is pushing much the RT cores....dunno if that is true for Octane itself as well...

guess 3090 will be 700-900 OB, really huge improvement, we may find it out in a week

jimho wrote:guess 3090 will be 700-900 OB, really huge improvement, we may find it out in a week

Jules posted on twitter a little while back and mentioned the 3090 is about 700 OB
https://twitter.com/otoy/status/1306950 ... 90529?s=21

So my theoretical calculation was right with 684, not taking into account the little higher base clock, which would lead to around 710 (o;

So pudgetsystems published the first result in OctaneBench with a score of 670.6....

https://www.pugetsystems.com/labs/artic ... ance-1890/

Close to my calculation (o;

Article here: https://techgage.com/article/nvidia-gef ... rformance/

davorin wrote:Stupid me (o;
Of course the RTX3080 can't be twice as fast as the RTX2080TI (o;
RTX3080 has twice the CUDA core amount, but the RT core amount is the same...
Therefore I can assume that doubling CUDA cores takes 60% of the speed improvement, whereas RT cores 40%.


So a RTX3090 with 10496 cores has 2.412 times more CUDA cores as RTX2080TI, and 60% of that gives a point increase of around 300 ( (10496 / 4352 * 355 - 355) * 0.6)...
And the RT core increase from 68 to 82 gives a point increase of around 29...( (82 / 68 * 355 - 355) * 0.4)

So my assumption is that the RTX3090 should be 684, or 1.93 times faster than a RTX2080TI...
but with the base clock being a little higher, it should be at least factor 2...

In another thread here I read that OctaneBench is pushing much the RT cores....dunno if that is true for Octane itself as well...

Not sure if it's already been mentioned elsewhere.
I don't think you can directly compare Ampere's CUDA core count with previous generations.
From what I've read in a couple of reviews, the previous generation had X number of CUDA cores, but that number represented two separate banks of floating-point-operation cores, and integer-operation cores.
Ampere has an entirely different structure - half of the apparently-huge number of cores are floating-point, the other half are both floating-point AND integer, and can be switched between doing either type of operation.
...so if you happen to have a workload that does absolutely zero integer maths, then yes, the Ampere cards will have a massive performance boost. But I assume that no such workload actually exists.
For games, nVidia themselves have estimated that typically around 65% of operations run are floating-point.
For GPU-accelerated rendering, I have no idea what the balance would typically be.

At the end of the day, all that's certain is there's pretty much no point attempting to directly compare this architecture with the previous one.
That probably goes for RT-core counts too, as I've read plenty of reports saying that they've significantly changed their design (although nobody seems to actually know in what way )