The GT300/Fermi Thread - Part 2!

[Teemax]

Regarding the tesselation performance of Fermi, I'd wait to see the comparison in real world games. IF (and I don't know if that's true, false, or somewhat in between) NV100 relies more in CUDA processors to solve tesselation calculations and RV870 relies more in fixed performance dedicated hardware, I'd expect the first to have a huge advantage in a nearly synthetic benchmark with most of its load being tesselation (because Fermi could use more resources to do it), but then, the situation would rebalance severely when complex shaders should be computed in addition to tesselation (like it would be the most likely case in most real world games). The heaven benchmark seems to be pretty heavy on tesselation, but much lighter in any other kind of shader. Maybe that's why NV are focusing so much on Unigine Heaven benchmark. Maybe not. We will see... when we have proper reviews and real world use cases.

Exactly my thoughts. I'm not buying NVIDIA's magical tessellation performance story yet.

I have doubt that NVIDIA simply invest so much in dedicated tessellation hardware. It's probably more likely that they use the "cuda cores" to do tessellation in exchange for shader performance.

If that's the case, their DX 11 driver will have some delicated load balancing to do. It must decide how many cores should be reserved for tessellation for each game. Not neccessarily an elegant solution.


Until we see performances of real games, I'm not holding my breath.

[BenchZowner]

I have doubt that NVIDIA simply invest so much in dedicated tessellation hardware. It's probably more likely that they use the "cuda cores" to do tessellation in exchange for shader performance.

If that's the case, their DX 11 driver will have some delicated load balancing to do. It must decide how many cores should be reserved for tessellation for each game. Not neccessarily an elegant solution.

If what you said is true, then nVIDIA lied in a public technical specifications pdf.
And you're wrong about load balancing via the driver.
That would be the silliest move ever.
"CUDA cores" can be dynamically assigned to various tasks and load balancing should and is done by hardware not software.

Not an elegant thinking I'd say

[saaya]

If what you said is true, then nVIDIA lied in a public technical specifications pdf.
And you're wrong about load balancing via the driver.
That would be the silliest move ever.
"CUDA cores" can be dynamically assigned to various tasks and load balancing should and is done by hardware not software.

Not an elegant thinking I'd say

well, when is something accelerated in hardware and when is it just being emulated? its hard to tell, and as soon as hardware detects a certain type of code or can switch into a different mode of operation which is faster for certain code, you could already call that hardware acceleration and dedicated hardware for that code... even if the same logic is actually able to process very different code... the propper definition for dedicated hardware or hardware accelration is to have some logic that is specific to a certain type of code and can ONLY process and accelerate that code and can not be used for anything else. im pretty sure thats what tesselation in gf100 is NOT... it simply would go against nvidias design goal of having a general purpose monster flop throughput processor...

so did nvidia lie? i wouldnt say so...
and even if they did, there is no way to prove them wrong, it will always be an argument of different interpretations and definitions... and in the end, like i said before... if it performs well, who cares?
will it actually perform well in a real world scenario is something that we are all curious about though... i cant wait for avp and some other actual game benchmarks with gf100...

in the end i dont think gf100 will be vastly superior in tesselation compared to rv870... it will be faster, yes, but the difference isnt 100% like nvidia first claimed... in a synthetic benchmark, the best case scenario, in a certain scene, so the best case of a best case, gf100 is double as fast as rv870... in reality the difference is probably below 50%, so 30fps vs 40fps in avp or something like that...

tesselation is NOT a killer feature for gf100 if you ask me...

[Teemax]

If what you said is true, then nVIDIA lied in a public technical specifications pdf.
And you're wrong about load balancing via the driver.
That would be the silliest move ever.
"CUDA cores" can be dynamically assigned to various tasks and load balancing should and is done by hardware not software.

Not an elegant thinking I'd say

Which part of the "specifications pdf" would be wrong if my claims are true?

NVIDIA claimed that GF100 has stellar tessellation performance, which might well be true. But they never claimed that tessallation wouldn't affect shader performance.

If you have info that disproves what I said, by all means do share it.

[Chumbucket843]

Exactly my thoughts. I'm not buying NVIDIA's magical tessellation performance story yet.

I have doubt that NVIDIA simply invest so much in dedicated tessellation hardware. It's probably more likely that they use the "cuda cores" to do tessellation in exchange for shader performance.

If that's the case, their DX 11 driver will have some delicated load balancing to do. It must decide how many cores should be reserved for tessellation for each game. Not neccessarily an elegant solution.


Until we see performances of real games, I'm not holding my breath.

its getting ridiculous how many people think gf100 does not have fixed function tessellation. if they are that incompetent maybe they should hire people off of tech forums to architect their gpu's. most people probably dont know the difference from fixed function logic or programmable logic anyways.

what they did was fairly simple. gf100 basically is setting up the scene in parallel compared to serial setup of other gpus. it works well for all of the small triangle tessellation creates.

[Iconyu]

its getting ridiculous how many people think gf100 does not have fixed function tessellation. if they are that incompetent maybe they should hire people off of tech forums to architect their gpu's. most people probably dont know the difference from fixed function logic or programmable logic anyways.

what they did was fairly simple. gf100 basically is setting up the scene in parallel compared to serial setup of other gpus. it works well for all of the small triangle tessellation creates.

The 5870 isn't incredibly fast on Dirt2 in DX11, playable yes but you're losing a lot of FPS. It wouldn't be that hard to prove the case eitherway, but in Dirt2 tessellation is done on top of everything else, not instead of something else, same with AvP.

I suppose in a few weeks we'll see the numbers but the fact that they are missing is worrying.