Nerfed FP64 performance in consumer GPU cards

[Sword7]

Folks,

I googled some FP64 (double precision) performance about any consumer GPU cards but learned that NVIDIA and AMD nerfed FP64 performance in consumer GPU cards. Also I learned that many people asked for FP64 hardware support in several years.

I am concerning about that because of jittery issues when rendering large worlds with FP32 operations like space flight simulators, etc. There are good FP32 performance but inaccurate (limited 7 digits accuracy).

Games need physics and 64-bit vertices for large world to be calculated through GPU process for better performance.

I researched deeper and deeper and learned that AMD have little better FP64 performance than NVIDIA (1/32 FP32 performance).

Does any GPU hardware vendors plan to improve FP64 performance through consumer GPU cards?

Thanks,
Tim

 

[Selaya]

NVIDIA RTX 6000 Ada Generation Graphics Card

Powered by the NVIDIA Ada Lovelace Architecture.

www.nvidia.com

NVIDIA TITAN V Specs

NVIDIA GV100, 1455 MHz, 5120 Cores, 320 TMUs, 96 ROPs, 12288 MB HBM2, 848 MHz, 3072 bit

www.techpowerup.com

 

[111frodon]

This is why the professionnal cards exists. Market segmentation.

 

[Vya Domus]

It's not nerfed, the units don't exist physically on the chip and they're just not really required for rendering or game physics, in fact FP16 is more useful probably. Remember that you could have an immense high precision polygon count but at the end it will get rasterized into a 2D image.

Vertex data simply does not require precision that big, in fact some engines employed 16 bit storage for vertex data in order to save space.

Even though it seems like FP64 performance is very low it's still more than what most CPUs can do. I wrote several GPU accelerated algorithms and they all ran faster using FP64 on my 1080 which has a 1/32 FP64 ratio.

 

[newtekie1]

It's not nerfed, the units don't exist physically on the chip and they're just not really required for rendering or game physics, in fact FP16 is more useful probably.

That's not always true. Yes, in the Ampere GPUs, that is the case I believe. The GA100 used in Ampere high end compute cards does have more FP64 units than the GA102 used in the RTX3090/3080.

However, in other cards the FP64 performance is just nerfed. The RTX3090 and Quadro A6000 both use the same GA102 GPU core. So the FP64 units in the cards is the same. However, the Quadro A6000's FP performance is 1:32 of the FP32 but the RTX3090 is 1:64. So there are just far more FP64 units disabled(but still physically there) in the RTX3090 or the performance is nerfed. And that's pretty much how it all goes in the consumer space vs the professional space. Cards that use the same GPU cores have far lower FP64 performance in consumer cards despite the FP64 units being present in the GPU.

Now it also could be entirely a yield thing too. Just like sharders are disabled because they are damaged and the GPU is used in a lower tier card the same might be done with FP64 units. If there are enough damaged FP64 units, then the GPU is downgraded for use in a consumer card.

 

[agent_x007]

Just buy old Titan if you need FP64 that badly, and be happy (after some OC).

 

[Sandbo]

Just buy old Titan if you need FP64 that badly, and be happy (after some OC).
View attachment 170178

Or if you can find one, Radeon VII has a less crippled FP64 performance at ~ 3 TFLOP.

 

[Lethos009]

Sorry if my post is misplaced, probably should start a new topic but this one is kinda exactly my problem.

Do you happen to know, is it possibile that fp64 perf was nerfed in some of the latest drivers? Or maybe some weird Aida64 quirks? Because I've recently bought an used GTX Titan (not black, which I know it should perform just slightly worse), thinking mostly about this fp64 performance, which is beneficial in my university work. I've ran the test, but the score for double-precision flops for my card was terrible. I'm not sure if that's right and I missed something about nerfing fp64, or maybe bugged drivers, or maybe I was scammed? Any of your help, I would really appreciate it

 

[damric]

Sorry if my post is misplaced, probably should start a new topic but this one is kinda exactly my problem.

Do you happen to know, is it possibile that fp64 perf was nerfed in some of the latest drivers? Or maybe some weird Aida64 quirks? Because I've recently bought an used GTX Titan (not black, which I know it should perform just slightly worse), thinking mostly about this fp64 performance, which is beneficial in my university work. I've ran the test, but the score for double-precision flops for my card was terrible. I'm not sure if that's right and I missed something about nerfing fp64, or maybe bugged drivers, or maybe I was scammed? Any of your help, I would really appreciate it

View attachment 174387View attachment 174388

You would be way better off with a used Vega.

 

[Lethos009]

That's the point, Titan should easily smoke your Vega in case of FP64 but somehow it does not. Apart from that, CUDA support is crucial for me, which unfortunately is not available on Vega.

 

[DR4G00N]

That's the point, Titan should easily smoke your Vega in case of FP64 but somehow it does not. Apart from that, CUDA support is crucial for me, which unfortunately is not available on Vega.

Yeah, somethings off there, a Titan should be around 1.3TFLOPS. Try an older driver and see if that changes anything.

 

[Rei]

Because I've recently bought an used GTX Titan (not black, which I know it should perform just slightly worse)

How would you know this? That is very unlikely & should be the opposite.

 

[Lethos009]

How would you know this? That is very unlikely & should be the opposite

Well, I wrote it wrongly. What I meant, is that I know my card should perform worse than Titan Black, but not like I've experienced. Sorry, english isn't my native language.

Anyway, I tried couple different driver versions, also different versions of Aida64 - still the same result Quickly swapped for my previous card, 750Ti - and the results were normal just like on the website. Soo I'm getting more and more concerned Would love to see some comparison from person who also have Titan/Titan Black, before going for Windows reinstall

 

[Rei]

Well, I wrote it wrongly. What I meant, is that I know my card should perform worse than Titan Black, but not like I've experienced. Sorry, english isn't my native language.

Ok, don't worry about it. I was just curious.

Anyway, I tried couple different driver versions, also different versions of Aida64 - still the same result Quickly swapped for my previous card, 750Ti - and the results were normal just like on the website. Soo I'm getting more and more concerned

a) Did you use the "generic" driver from Nvidia's website?
b) What brand is your Titan from?
c) Have you tried to use the driver from your Titan's brand's website?

Would love to see some comparison from person who also have Titan/Titan Black, before going for Windows reinstall

Sorry, I only have GTX 780 Ti which has nearly the same spec as Titan Black but far inferior FP64 of only 0.21 TFLOPS compared to Titan Black's 1.7 TFLOPS.

 

[silentbogo]

Sorry if my post is misplaced, probably should start a new topic but this one is kinda exactly my problem.

Do you happen to know, is it possibile that fp64 perf was nerfed in some of the latest drivers? Or maybe some weird Aida64 quirks? Because I've recently bought an used GTX Titan (not black, which I know it should perform just slightly worse), thinking mostly about this fp64 performance, which is beneficial in my university work. I've ran the test, but the score for double-precision flops for my card was terrible. I'm not sure if that's right and I missed something about nerfing fp64, or maybe bugged drivers, or maybe I was scammed? Any of your help, I would really appreciate it

Just a quick google-foo found a solution for your problem. FP64 acceleration by default is disabled, because it gimps FP32 performance. Basically what you are getting is a 1/24 ratio, like on GTX780. If you enable FP64, you'll get your 1:3 ratio, but expect a slight performance drop in FP32.

 

[Lethos009]

c) Have you tried to use the driver from your Titan's brand's website?

No I haven't, I had no idea that such website even exists

Just a quick google-foo found a solution for your problem. FP64 acceleration by default is disabled, because it gimps FP32 performance. Basically what you are getting is a 1/24 ratio, like on GTX780. If you enable FP64, you'll get your 1:3 ratio, but expect a slight performance drop in FP32.

That's great! Do you know how to switch it on? As far as I'm concerned, I need as much FP64 as it gets me, I don't care gaming performance. Also, if it's as simple as turning it on/off, I could test if it's beneficial for me

 

[silentbogo]

Open NVidia Control Panel, go to "Manage 3D Settings", scroll down a bit and you'll see it as "Double Precision" option.
By default it's set to None. You need to switch it to your Titan.

 

[Lethos009]

Open NVidia Control Panel, go to "Manage 3D Settings", scroll down a bit and you'll see it as "Double Precision" option.
By default it's set to None. You need to switch it to your Titan.

Yep! That's it. Huge thanks to you man, I was getting so worried.. Thank you so much!

 

[agent_x007]

Keep in mind this will decrease max clock on the card hard in games, and probably make it run hotter/louder on top (because this setting makes hard offset on vGPU, at least that's what it does on my Titan Black).