why does gtx 780 have bad integer performance?

[Anarchist420]

the newest build of dolphin runs slow as hell on my GTX 780. is the titan any better in that regard?

 

[Techhog]

Are you sure that you don't have a CPU problem?

 

[Anarchist420]

Are you sure that you don't have a CPU problem?

yes, because older builds were pretty fast

 

[Bubbleawsome]

Go back to old versions?

 

[ShintaiDK]

Your problem seems to be your software version and not the card.

If it worked in a previous version, its obviously not the card.

 

[Spjut]

Are you perhaps using OpenGL? Dolphin's blog mentions specifically that Nvidia only takes a 5% hit compared to before using DirectX, but "significant losses" if using OpenGL

https://dolphin-emu.org/blog/2014/03/15/pixel-processing-problems/

 

[Anarchist420]

Are you perhaps using OpenGL? Dolphin's blog mentions specifically that Nvidia only takes a 5% hit compared to before using DirectX, but "significant losses" if using OpenGL https://dolphin-emu.org/blog/2014/03...sing-problems/

dx backend sucks because it doesnt have GL backend's aa mode that i like. why shouldnt everyone be able to use opengl and still get full performance? he mentions that nv's integer performance is weak except on what i have and the titan but my performance suffers anyway. i am wondering if titan has integer performance on par with AMD's products

 

[TreVader]

because it takes 3x as long to do an integer rotate on Kepler vs Tahiti/hawaii

 

[Anarchist420]

because it takes 3x as long to do an integer rotate on Kepler vs Tahiti/hawaii

well that sucks balls because everything else on the GK110 is better (and i dont know why nvidia's engineers thought that integer performance was unimportant because a lot of legacy functions used integers and single precision float gives too much error when integer legacy functions are emulated). amd doesnt even do full speed FP16 TF, they dont have physx or CUDA, they dont have much driver-forced aa, and i dont think their TF units can be forced to do trilinear mipmapping regardless of application nor has AMD ever had very good trilinear calculations (unless that actually changed from the 6k series to the 7k series).
AMD reference coolers arent quite as good and AMD chips run at high temps.

one thing nv has started with all drivers compatible with kepler and that AMD always did that i am not to happy about is some depth range optimization when driver SGSSAA isnt forced.

yet AMD can have better integer performance?

sorry about the rant.

 

[Atreidin]

With chips so complex, of the same generation, so close in general performance, but still pretty different designs, you will always find something one of them is better at than the other, even if it is just a few instructions that one runs in less clock cycles, but most programs don't use much. Even if one is totally creaming the other in most consumer applications you could probably pore over the datasheets and find instructions the "faster" one is worse at than the "slower" one. This is true of GPUs, CPUs, whatever.

 

[x3sphere]

dx backend sucks because it doesnt have GL backend's aa mode that i like. why shouldnt everyone be able to use opengl and still get full performance? he mentions that nv's integer performance is weak except on what i have and the titan but my performance suffers anyway. i am wondering if titan has integer performance on par with AMD's products

it doesn't. Maxwell (750 Ti) has better integer performance though I believe, which bodes well for the upcoming high end cards.

 

[Techhog]

well that sucks balls because everything else on the GK110 is better (and i dont know why nvidia's engineers thought that integer performance was unimportant because a lot of legacy functions used integers and single precision float gives too much error when integer legacy functions are emulated). amd doesnt even do full speed FP16 TF, they dont have physx or CUDA, they dont have much driver-forced aa, and i dont think their TF units can be forced to do trilinear mipmapping regardless of application nor has AMD ever had very good trilinear calculations (unless that actually changed from the 6k series to the 7k series).
AMD reference coolers arent quite as good and AMD chips run at high temps.

one thing nv has started with all drivers compatible with kepler and that AMD always did that i am not to happy about is some depth range optimization when driver SGSSAA isnt forced.

yet AMD can have better integer performance?

sorry about the rant.

So, basically, you're mad because GK110 doesn't win in every single way. Huh.

 

[TreVader]

well that sucks balls because everything else on the GK110 is better (and i dont know why nvidia's engineers thought that integer performance was unimportant because a lot of legacy functions used integers and single precision float gives too much error when integer legacy functions are emulated). amd doesnt even do full speed FP16 TF, they dont have physx or CUDA, they dont have much driver-forced aa, and i dont think their TF units can be forced to do trilinear mipmapping regardless of application nor has AMD ever had very good trilinear calculations (unless that actually changed from the 6k series to the 7k series).
AMD reference coolers arent quite as good and AMD chips run at high temps.

one thing nv has started with all drivers compatible with kepler and that AMD always did that i am not to happy about is some depth range optimization when driver SGSSAA isnt forced.

yet AMD can have better integer performance?

sorry about the rant.

What?



Kepler is great but so is Hawaii. If AMD wasn't around Nvidia would be working a lot less hard on making good products. Look at intel now that AMD isn't big competition.

 

[Anarchist420]

So, basically, you're mad because GK110 doesn't win in every single way. Huh.

well not really because i cant afford a non-defective GK110 anyway (i.e., one with good DP performance). thank you

What? Kepler is great but so is Hawaii. If AMD wasn't around Nvidia would be working a lot less hard on making good products. Look at intel now that AMD isn't big competition.

i guess youve a good pt

it doesn't. Maxwell (750 Ti) has better integer performance though I believe, which bodes well for the upcoming high end cards.

it has even worse DP performance though. really advanced physics simulations could have too much error with only SP. particle fx and fog effects could have too much error with SP if they are really complex. i just think it is time to do everything except TA/TF units, and display logic in software with double precision (for float and int). larrabee was too slow because of the isa it used. i realize that the current CUDA architecture would be too slow, but that doesnt mean that a new gen purp ISA (with a bias towards graphics) cant be created.

one other thing i dont understand is why GPU dies and memory arent removable and replaceable. they could use fluxless solder to reduce GPU die temps if the pcbs had sockets

 

[Silverforce11]

it has even worse DP performance though. really advanced physics simulations could have too much error with only SP. particle fx and fog effects could have too much error with SP if they are really complex. i just think it is time to do everything except TA/TF units, and display logic in software with double precision (for float and int).

There already exists a focused product that caters to your needs, Teslas and Quadros. Gamers have no need for those features so getting streamlined, albeit DP neutered but efficient gaming chips is a better deal for gamers, in terms of perf/w and perf/$.

 

[BFG10K]

yes, because older builds were pretty fast

Then it's not a hardware problem.

why shouldnt everyone be able to use opengl and still get full performance?

How would we know? Why don't you ask the developer instead?

You know, the guy that actually wrote the code.

 

[Anarchist420]

How would we know? Why don't you ask the developer instead? You know, the guy that actually wrote the code.

i could try doing that, but i doubt they will respond. i might try asking them though.

There already exists a focused product that caters to your needs, Teslas and Quadros. Gamers have no need for those features so getting streamlined, albeit DP neutered but efficient gaming chips is a better deal for gamers, in terms of perf/w and perf/$.

gamers have no need for those features only because they arent available on mainstream cards. that is, game devs wont use them or will be less likely to use them because only a small percentage of people buy $1k GPUs.

not that i think AMD's products are very good overall, but their non-workstation products have better compute features than nv's non-ws do which must mean that extensive compute features could be useful for games.

 

[tweakboy]

Don't worry about it or scores or benchies. Just use your CPU and do something productive on it.

 

[FalseChristian]

Maxwell will have much higher Compute performance in the non-Titan cards so don't worry. Be happy.

 

[lamedude]

Here's what Dolphin's NV contact said.

Our GPUs are notoriously "not great" at integer math. [...] It's also possible that the D3D driver (which uses much more aggressive settings in the compiler and continues to optimize in the background when GPU limited) figured out that you are doing math that could be mostly done in float and converted back just in time.

 

[Anarchist420]

Here's what Dolphin's NV contact said.

thank you lamedude