Nvidia 3070Ti - 21.75 TFlops FP32, ~$550 ebay price at present (recently sold listings)
AMD 6900XT - 23.04TFlops FP32, ~$600 ebay price at present (recently sold listings)
3070Ti FP32/$ = 0.0395 TF/$
6900XT FP32/$ = 0.0384 TF/$
seems pretty close right? even with the slight edge to Nvidia on FP32/$ these should be comparable performance. but in reality with the best applications currently available here, the 3070Ti is significantly faster. my 200W 3070Ti does each task in about 100 seconds (2x in 200s). whereas the top 6900XT system on the leaderboard is doing 4x in ~575s (143s per task), making the 3070Ti in reality ~40-45% faster for FGRPB1G.
Some projects do better on one, some better on the other, probably depends how optimised the Cuda code is. I detest Nvidia for throttling OpenCL.
As for the prices, I will not spend more on a GPU than my car! I've had a $600 car for 5 years (40K miles on the clock when I bought it, 95K now). I never spend over $300 on a computer part. When you go down to that price range, AMD is vastly cheaper.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
I’m talking about the OpenCL FGRPB1G apps, which is FP32 and directly applicable to Einstein. The Nvidia one is still OpenCL. Nvidia doesn’t throttle OpenCL, it’s just that there is a translation from OpenCL->CUDA that’s necessary because Nvidia hardware only runs CUDA natively. OpenCL apps are translated on the fly at runtime. There will always be some overhead, but it’s not nearly as bad as it used to be. The problem is that many devs don’t have the knowledge or time to optimize the OpenCL code for Nvidia.
put older/cheaper cards against each other, like the RX 580 vs the GTX 1060 against each other and you get much the same result.
Well I don't think it's that big a difference It depends how much you're paying, 2nd hand or new, what project you run, probably cost more in different countries. But I've never found Nvidia to be on top. It's good we have competition though. Same goes for the CPUs. Nothing gets improved if they don't have to. I always buy 280X cards. $80 2nd hand. Last about 2 years then explode. But if they explode in different ways you can put two together and get another couple of years.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
I'm okay with what your doing and believe in. I'm even okay with your saying that AMD is superior to NVIDIA in FP32 output per $$. But I'm not okay with your trying to convince me and the rest of us on BOINC that AMD is better. Can't you let the topic rest?
I will continue to buy NVIDIA, not because of the performance per $$ when comparing FP32, but because I like the peace of mind that NVIDIA gives me in their software updates (drivers) and by purchasing from EVGA with superior product support.
The support is the same, mostly they both work, occasionally there are glitches. The only way to choose is FP32/$.
Peter Hucker of the Scottish Boinc Team wrote:
..... "The ONLY way to choose is FP32/$" .....
?????
This website happens to have some pretty good statements worth taking note of, like:
"The FP32 raw performance is of a purely theoretical nature, as it is only a small part of a complex GPU."
It essentially tells me what key features at a glance by each comparable GPU. Then I make my decision to purchase based upon price and the peace of mind that NVIDIA gives me in their software updates (drivers) and by purchasing from EVGA with superior product support, as compared to JUST FP32/$$.
I'm not expecting you to agree with me, only to let me and others have a say without you trying impress upon us that your way is THE ONLY WAY. It is the way that YOU do it. Not the way that WE do it.
It essentially tells me what key features at a glance by each comparable GPU. Then I make my decision to purchase based upon price and the peace of mind that NVIDIA gives me in their software updates (drivers) and by purchasing from EVGA with superior product support, as compared to JUST FP32/$$.
I'm not expecting you to agree with me, only to let me and others have a say without you trying impress upon us that your way is THE ONLY WAY. It is the way that YOU do it. Not the way that WE do it.
Simplest thing to do is speed versus cost. I pick a car the same way. If they've made it go fast, they've also designed it well. I'm not going to start delving into instructions sets and what's best for particular projects, I might be running another project in a few months time.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
[sulks] I want AMD on Windows. Why did you have to program the niche stuff first?
The Windows OpenCL version (for AMD and NVidia) was already tested successfully. And btw, the "niche" is rather AMD, see the sever status page. E@H has 2-3x as many hosts with NVidia cards than with AMD. And I think having more than one card per host is more common with NVidia than AMD, so the card/GPU count will be even more in favor for NVidia.
Are there any Meerkat tasks for ARM cpu's or you guys think they will take waaaay too long?
Peter, I'm not saying that I'm right and you are wrong. It is just that this type of conversation doesn't belong in "Technical News from this project". It has no bearing on "Technical News" whatsoever.
It doesn't address the latest cards at all, but I do believe it has merit for older cards, not unlike yours.
Regardless...
Please take the conversation(s) expressing YOUR particular views somewhere else, like "Cafe Einstein - Discuss anything relating to Einstein@home". That's a good place to put your ideas and preferences to work for everyone and anyone to read, start your own thread on the topic.
[sulks] I want AMD on Windows. Why did you have to program the niche stuff first?
The Windows OpenCL version (for AMD and NVidia) was already tested successfully. And btw, the "niche" is rather AMD, see the sever status page. E@H has 2-3x as many hosts with NVidia cards than with AMD. And I think having more than one card per host is more common with NVidia than AMD, so the card/GPU count will be even more in favor for NVidia.
Are there any Meerkat tasks for ARM cpu's or you guys think they will take waaaay too long?
based on the messages posted early in the thread, my impression is that the BRP7 MeerKAT tasks will be for GPUs only.
there is still plenty of BRP4 work for slower CPUs and ARM devices to crunch on.
Theoretical vs
)
Theoretical vs Reality
Nvidia 3070Ti - 21.75 TFlops FP32, ~$550 ebay price at present (recently sold listings)
AMD 6900XT - 23.04TFlops FP32, ~$600 ebay price at present (recently sold listings)
3070Ti FP32/$ = 0.0395 TF/$
6900XT FP32/$ = 0.0384 TF/$
seems pretty close right? even with the slight edge to Nvidia on FP32/$ these should be comparable performance. but in reality with the best applications currently available here, the 3070Ti is significantly faster. my 200W 3070Ti does each task in about 100 seconds (2x in 200s). whereas the top 6900XT system on the leaderboard is doing 4x in ~575s (143s per task), making the 3070Ti in reality ~40-45% faster for FGRPB1G.
_________________________________________________________________________
Some projects do better on
)
Some projects do better on one, some better on the other, probably depends how optimised the Cuda code is. I detest Nvidia for throttling OpenCL.
As for the prices, I will not spend more on a GPU than my car! I've had a $600 car for 5 years (40K miles on the clock when I bought it, 95K now). I never spend over $300 on a computer part. When you go down to that price range, AMD is vastly cheaper.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
I’m talking about the OpenCL
)
I’m talking about the OpenCL FGRPB1G apps, which is FP32 and directly applicable to Einstein. The Nvidia one is still OpenCL. Nvidia doesn’t throttle OpenCL, it’s just that there is a translation from OpenCL->CUDA that’s necessary because Nvidia hardware only runs CUDA natively. OpenCL apps are translated on the fly at runtime. There will always be some overhead, but it’s not nearly as bad as it used to be. The problem is that many devs don’t have the knowledge or time to optimize the OpenCL code for Nvidia.
put older/cheaper cards against each other, like the RX 580 vs the GTX 1060 against each other and you get much the same result.
_________________________________________________________________________
Well I don't think it's that
)
Well I don't think it's that big a difference It depends how much you're paying, 2nd hand or new, what project you run, probably cost more in different countries. But I've never found Nvidia to be on top. It's good we have competition though. Same goes for the CPUs. Nothing gets improved if they don't have to. I always buy 280X cards. $80 2nd hand. Last about 2 years then explode. But if they explode in different ways you can put two together and get another couple of years.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Peter Hucker of the Scottish
)
?????
This website happens to have some pretty good statements worth taking note of, like:
"The FP32 raw performance is of a purely theoretical nature, as it is only a small part of a complex GPU."
https://gpu-benchmark.com/benchmarks/fp32-performance-single-precision-tflops/
But I personally like to make comparisons between GPUs using websites such as Passmark's Price Performance.
https://www.videocardbenchmark.net/high_end_gpus.html
It essentially tells me what key features at a glance by each comparable GPU. Then I make my decision to purchase based upon price and the peace of mind that NVIDIA gives me in their software updates (drivers) and by purchasing from EVGA with superior product support, as compared to JUST FP32/$$.
I'm not expecting you to agree with me, only to let me and others have a say without you trying impress upon us that your way is THE ONLY WAY. It is the way that YOU do it. Not the way that WE do it.
Proud member of the Old Farts Association
GWGeorge007
)
Simplest thing to do is speed versus cost. I pick a car the same way. If they've made it go fast, they've also designed it well. I'm not going to start delving into instructions sets and what's best for particular projects, I might be running another project in a few months time.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Bernd Machenschalk
)
Are there any Meerkat tasks for ARM cpu's or you guys think they will take waaaay too long?
Peter, I'm not saying that
)
Peter, I'm not saying that I'm right and you are wrong. It is just that this type of conversation doesn't belong in "Technical News from this project". It has no bearing on "Technical News" whatsoever.
It doesn't address the latest cards at all, but I do believe it has merit for older cards, not unlike yours.
Regardless...
Please take the conversation(s) expressing YOUR particular views somewhere else, like "Cafe Einstein - Discuss anything relating to Einstein@home". That's a good place to put your ideas and preferences to work for everyone and anyone to read, start your own thread on the topic.
Proud member of the Old Farts Association
GWGeorge007 wrote: anyone to
)
Here, Here!
Peter,
I promise to subscribe to a Peter thread on gpus and other topics. Even agitate you gently so you won't feel like you are in an silent echo chamber ;)
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
mikey wrote: Bernd
)
based on the messages posted early in the thread, my impression is that the BRP7 MeerKAT tasks will be for GPUs only.
there is still plenty of BRP4 work for slower CPUs and ARM devices to crunch on.
_________________________________________________________________________