CUDA cores per frame

[Not I Said the Cat]

Let's say that you have benchmarked a game using the GTX 1050. Assuming no memory bottlenecks, could you then determine the fps of the game using a higher end GPU. In other words, can the number of CUDA cores determine the framerate of games, and how would you adjust for frequency? Just curious.

[Silentprototipe]

Thats now how it works im affraid....

[Bananasplit_00]

if we are talking pascal only, kind of. its not a linial scale though and clock speeds need to be kept in mind

[RadiatingLight]

Yes, but with a lot of qualifiers

 

Firstoff, CUDA cores (and clockspeed) can only be compared/counted/etc. within the same generation

This is extremely important. a GTX 980Ti has more CUDA cores than a GTX 1080, but it's worse. this is because you cannot compare CUDA core number, clockspeed, etc. between generations (Maxwell and pascal in this case)

obviously, CUDA cores and AMD Stream processors can't be compared under any circumstances.

 

Secondly, CUDA cores aren't everything.

GPUs are more than just CUDA cores / Stream processors. they have ROPs (Render Output Units), VRAM, internal connections, etc. Unless you take all these into account (really hard), you won't get an exact estimate. if you're counting CUDA cores, you'll get an approximation, which is usually good enough.

 

Third, Clockspeed can be just added together.

Clockspeed is the simplest. if you increase the clockspeed of your GPU core to 10Ghz from 2Ghz, then you'll theoretically see a 5X increase in performance. however, you also need to take into account the other things on the GPU (ROPs, VRAM, etc.). if you increased the clockspeed of all of those things by 5X, then you would see 5X the performance. for example, a Ryzen 1600 brought from 3Ghz to 4Ghz would yield a theoretical improvement of 33%, since 4 is 33% more than 3.

But, clock speed can only be added within the same architecture/generation, and company. just like CUDA cores.

keep in mind that this is pretty theoretical.

 

Finally, Nothing here is exact -- just look at benchmarks

there are tons of variables if you want to theoretically calculate the performance of a GPU, and if you're just looking for a rough performance estimate, they might work. but there are much better ways of seeing the performance of a GPU. just look at benchmarks. they'll be more accurate, and easier to find than calculating the CUDA core difference.

[Enderman]

There is a lot more that determines performance than just cores, so no, not at all.

[Not I Said the Cat]

On 11/18/2017 at 7:31 PM, RadiatingLight said:

Yes, but with a lot of qualifiers

 

Firstoff, CUDA cores (and clockspeed) can only be compared/counted/etc. within the same generation

This is extremely important. a GTX 980Ti has more CUDA cores than a GTX 1080, but it's worse. this is because you cannot compare CUDA core number, clockspeed, etc. between generations (Maxwell and pascal in this case)

obviously, CUDA cores and AMD Stream processors can't be compared under any circumstances.

 

Secondly, CUDA cores aren't everything.

GPUs are more than just CUDA cores / Stream processors. they have ROPs (Render Output Units), VRAM, internal connections, etc. Unless you take all these into account (really hard), you won't get an exact estimate. if you're counting CUDA cores, you'll get an approximation, which is usually good enough.

 

Third, Clockspeed can be just added together.

Clockspeed is the simplest. if you increase the clockspeed of your GPU core to 10Ghz from 2Ghz, then you'll theoretically see a 5X increase in performance. however, you also need to take into account the other things on the GPU (ROPs, VRAM, etc.). if you increased the clockspeed of all of those things by 5X, then you would see 5X the performance. for example, a Ryzen 1600 brought from 3Ghz to 4Ghz would yield a theoretical improvement of 33%, since 4 is 33% more than 3.

But, clock speed can only be added within the same architecture/generation, and company. just like CUDA cores.

keep in mind that this is pretty theoretical.

 

Finally, Nothing here is exact -- just look at benchmarks

there are tons of variables if you want to theoretically calculate the performance of a GPU, and if you're just looking for a rough performance estimate, they might work. but there are much better ways of seeing the performance of a GPU. just look at benchmarks. they'll be more accurate, and easier to find than calculating the CUDA core difference.

Thank You, RadiatingLight

 

In response to your first point, I only meant within Pascal when I wrote my question. I also meant clock speed when I said frequency. I'll definitely be more careful to specify things like that in future posts. Overall, I'm thoroughly impressed with and very thankful for the depth and detail of your response. I didn't expect CUDA to be the only factor that would determine gpu performance, but I was just curious, and I am very surprised by the etiquette of this forum. I definitely will come here before anywhere else with my future questions.