I've been seeing
fanboys well intentioned raytracing enthusiast, cream themselves over the kepler Tweet about Ray triangle intersection. Funny enough a lot of them haven't realised that it doesn't actually show what they think it shows.
The conclusions you can draw from that chart are as follows
1. Nvidia design is less efficient with ray triangle intersection than AMDs.
2. AMD's design may scale better than Nvidia when increasing ray triangle intersection.
3. Let us assume that AMD design has hit the limit and Nvidia design is the only way forward. This is bad news for anyone who is a fan of RT.
And I will explain why that is right after this message from our sponsor RAID SHADOW LEGENDS!!!
1. Nvidia design is less efficient with ray triangle intersection than AMDs
Lets start from the top. What did certain people say about RDNA 3 RT ability.
THEY ARE A GENERATION BEHIND!!! LOL!!! THEY ARE ONLY AS FAST AS A 3090TI in RT!!!!
So AMD's new card is as fast as a 3090ti while only being able to process ~30% fewer ray triangle intersection per second.
Let that sink in. I'll wait.
2. AMD's design may scale better than Nvidia when increasing ray triangle intersection.
As someone on Twitter pointed out the 4090ti has 4x the ray triangle intersection operation per second than the 3090ti and is only twice as fast in actual games and 3D rendering.
What people didn't notice is that the 7900XTX has 16% more ray triangle intersection operation per second than the 6950XT, but has about double the performance from early indications.
You explain to me how that works.
3. Let us assume that AMD design has hit the limit and Nvidia design is the only way forward. This is very bad news for anyone who is a fan of RT.
Let us say that AMD design is a dead end and only Nvidia's design is the way to go.
From the 2080ti to the 3090ti, there is a 3x ray intersection increase for a doubling of performance(is this true?).
From the 3090ti to the 4090, there is a 4x ray intersection increase for a doubling of performance.
So for the 4090 to the 5090 are we looking at a 5x increase for a doubling of performance?
5090 to 6090, will that be a 6x increase for double the performance?
What about after that? How long can such a trend be maintained, especially with the slow down in node shrinkage?
That assumes that it is a linear increase, God help us if it is an exponential increase.
I think that tweet from Kepler was very informative. Just not for the reason some people thought