One might legitimately ask : what is the memory being allocated for? I think the answer is straightforward : for Discrete Fourier Transforms, which is the bulk of the effort of any work unit here at E@H. Essentially manipulating quite large arrays of data ..... as usual we convolve the data with some template in the time domain, thus multiplying their respective Fourier Transforms in the frequency domain.
Hence we can ask : why does that change suddenly at DF = 0.35 and 0.70 ? What is significant as I see it, for a given multiplicity, is that the memory jump from 0.70 to 0.75 is close to twice the memory jump from 0.35 to 0.40 ( also 0.70 is twice 0.35 ). This I reckon signals some sort of choice in the nature of operations like, for example, operand length ( specifically a doubling of the number of significant digits in the operands ). This is a stab in the dark of course.
In any case the graph demonstrates why a 'lesser' card might be able to do low DF units but go astray with higher DFs.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
It is "well-known" (or not) that multi-GPU setups prefer the blower-style of cooler i.e. 1 centrifugal fan sucking in air, blowing it through the GPU and air exiting outside the case. While each card may not reach the peak performance of the custom radial-fan coolers, overall the system will run cooler and cards won't cause thermal throttling in other cards (overall better performance):
This also applies for single-GPU system where the case has poor or restricted airflow e.g. HTPC. On my own machine (an upgraded OEM PC), a 2X radial-fan rx570 causes noticeable increase in CPU temps when running boinc GPU tasks. I had to add 1 more fan for static pressure to prevent thermal throttling of the CPU (just runs hotter but doesn't hit limit). The air blowing from the GPU inside the case can increase temp of other components, this is why gaming machines are usually configured with at least 2-3 case fans (many cases support up to 8 or more). But, it is not enough to compensate for multi-GPU's.
The radial fan designs are really only for single-GPU systems (for gamers) where multi-GPU doesn't help that use case much (because vendors aren't really supporting it for gaming although multiple GPUs were paradoxically actually working better in DX12 than the previous attempts until it was dropped). For the gamer, getting max performance from 1 GPU is the target (prefer radial fans).
The Reader may notice that virtually all professional graphics cards are blower-style. Rendering, computation, etc. ("real" work) prefer blowers (for above reasons). AMD's consumer-level cards will apparently not have blower reference designs (custom board almost never have blower designs) since the primary market is gaming.
Here is a blower type R5700 that I am running 3 of on a mining rack. Previously 2 were running in an ATX Cube. The open box price was right over $300 at that time.
VisionTek Radeon Graphic Card - 8 GB GDDR6 - 1.47 GHz Core - 256 bit Bus 901281
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
It is "well-known" (or not) that multi-GPU setups prefer the blower-style of cooler i.e. 1 centrifugal fan sucking in air, blowing it through the GPU and air exiting outside the case. While each card may not reach the peak performance of the custom radial-fan coolers, overall the system will run cooler and cards won't cause thermal throttling in other cards (overall better performance):
This also applies for single-GPU system where the case has poor or restricted airflow e.g. HTPC. On my own machine (an upgraded OEM PC), a 2X radial-fan rx570 causes noticeable increase in CPU temps when running boinc GPU tasks. I had to add 1 more fan for static pressure to prevent thermal throttling of the CPU (just runs hotter but doesn't hit limit). The air blowing from the GPU inside the case can increase temp of other components, this is why gaming machines are usually configured with at least 2-3 case fans (many cases support up to 8 or more). But, it is not enough to compensate for multi-GPU's.
The radial fan designs are really only for single-GPU systems (for gamers) where multi-GPU doesn't help that use case much (because vendors aren't really supporting it for gaming although multiple GPUs were paradoxically actually working better in DX12 than the previous attempts until it was dropped). For the gamer, getting max performance from 1 GPU is the target (prefer radial fans).
The Reader may notice that virtually all professional graphics cards are blower-style. Rendering, computation, etc. ("real" work) prefer blowers (for above reasons). AMD's consumer-level cards will apparently not have blower reference designs (custom board almost never have blower designs) since the primary market is gaming.
take this with the caveat that it depends a lot on what kind of setup you have.
my 7x and 10x GPU setups use dual fan cards, and perform better (lower temps, faster clock speeds) than they otherwise would with blower model cards. but these systems are not enclosed in a case, but instead on a mining style test bench.
the blower models work well when used in a server style case with high pressure fans, especially if they are right next to each other as the small gap between the cards does not afford good airflow from the suction of the GPU fan alone, it really needs the extra pressure from the chassis fans. i like the density of this kind of server setup, but the hardware to get it done for more than 2-3 GPUs in a single box gets into the realm of proprietary and expensive. spacing more normal hardware out on a mining style setup is much more economical. really the main reason for blowers on "professional" cards is because they are targetted for this kind of server environment with front to back airflow and high pressure fans. also most pro cards have lower clock speeds than their gaming counterparts, making them slightly slower for applications that wont take advantage of the other "pro" features that might exist on such cards, like massive amounts of GPU memory, ECC memory, or front facing power connectors (as opposed to top facing).
i kind of went half way on one system, in which i wanted to get maximum density and ran 7x GPUs plugged directly to the motherboard, but needed to use watercooling to shrink the size to single slot and an external radiator to dump the heat outside the room.
After re-reading the "Power Consumption" thread I poked around in the Radeon software I had, found "automatic tuning" with an "under voting" profile.
Now my R5700's are not blowing quite as hot as they were before. And I don't think I am seeing any lost performance (slower processing) either.
That's a win for my power bill and for my A/C trying to keep my house cool during the summer heatwave.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
When the next Gen of Navi cards comes out should we re-name this thread? Or start up a "Navi 10+ / Big Navi" thread?
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
I think I may have the highest ranked Navi 10 GPU (R5700's) now.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
I need to know what version of Win10 drivers for R5700's (Amd) are successful?
I am getting a steady stream of computational errors as well as invalids on my Win10 4 gpu box.
Tom M
-----Edit-------------
Never mind. It appears I was having a hardware error/malfunctioning gpu. I think my MB is not happy driving 4 gpus :(
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
I need to know what version of Win10 drivers for R5700's (Amd) are successful?
I am getting a steady stream of computational errors as well as invalids on my Win10 4 gpu box.
Tom M
-----Edit-------------
Never mind. It appears I was having a hardware error/malfunctioning gpu. I think my MB is not happy driving 4 gpus :(
Tom M
While running 3 gpus on this box ( https://einsteinathome.org/host/12835716 ) I am getting a small number of invalids.
The last one had this message in the output file.
"...
% Writing follow-up output file.
FPU status flags: COND_1 PRECISION
10:20:34 (7020): [normal]: done. calling boinc_finish(0).
10:20:34 (7020): called boinc_finish..."
And yesterday I was getting a lot more of the invalids
"...
Writing follow-up output file.
FPU status flags: PRECISION
13:53:59 (7344): [normal]: done. calling boinc_finish(0).
13:53:59 (7344): called boinc_finish
</stderr_txt>
]]>..."
Any ideas?
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
That's an interesting set of
)
That's an interesting set of curves.
One might legitimately ask : what is the memory being allocated for? I think the answer is straightforward : for Discrete Fourier Transforms, which is the bulk of the effort of any work unit here at E@H. Essentially manipulating quite large arrays of data ..... as usual we convolve the data with some template in the time domain, thus multiplying their respective Fourier Transforms in the frequency domain.
Hence we can ask : why does that change suddenly at DF = 0.35 and 0.70 ? What is significant as I see it, for a given multiplicity, is that the memory jump from 0.70 to 0.75 is close to twice the memory jump from 0.35 to 0.40 ( also 0.70 is twice 0.35 ). This I reckon signals some sort of choice in the nature of operations like, for example, operand length ( specifically a doubling of the number of significant digits in the operands ). This is a stab in the dark of course.
In any case the graph demonstrates why a 'lesser' card might be able to do low DF units but go astray with higher DFs.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
It is "well-known" (or not)
)
It is "well-known" (or not) that multi-GPU setups prefer the blower-style of cooler i.e. 1 centrifugal fan sucking in air, blowing it through the GPU and air exiting outside the case. While each card may not reach the peak performance of the custom radial-fan coolers, overall the system will run cooler and cards won't cause thermal throttling in other cards (overall better performance):
Explanations/details:
https://towardsdatascience.com/a-lesson-buying-a-multi-gpu-system-a14d3d1c710e
https://www.pugetsystems.com/blog/2019/01/11/NVIDIA-RTX-Graphics-Card-Cooling-Issues-1326/
https://www.pugetsystems.com/labs/articles/NVIDIA-Dual-Fan-GeForce-RTX-Coolers-Ruining-Multi-GPU-Performance-1239/
This also applies for single-GPU system where the case has poor or restricted airflow e.g. HTPC. On my own machine (an upgraded OEM PC), a 2X radial-fan rx570 causes noticeable increase in CPU temps when running boinc GPU tasks. I had to add 1 more fan for static pressure to prevent thermal throttling of the CPU (just runs hotter but doesn't hit limit). The air blowing from the GPU inside the case can increase temp of other components, this is why gaming machines are usually configured with at least 2-3 case fans (many cases support up to 8 or more). But, it is not enough to compensate for multi-GPU's.
The radial fan designs are really only for single-GPU systems (for gamers) where multi-GPU doesn't help that use case much (because vendors aren't really supporting it for gaming although multiple GPUs were paradoxically actually working better in DX12 than the previous attempts until it was dropped). For the gamer, getting max performance from 1 GPU is the target (prefer radial fans).
The Reader may notice that virtually all professional graphics cards are blower-style. Rendering, computation, etc. ("real" work) prefer blowers (for above reasons). AMD's consumer-level cards will apparently not have blower reference designs (custom board almost never have blower designs) since the primary market is gaming.
Here is a blower type R5700
)
Here is a blower type R5700 that I am running 3 of on a mining rack. Previously 2 were running in an ATX Cube. The open box price was right over $300 at that time.
VisionTek Radeon Graphic Card - 8 GB GDDR6 - 1.47 GHz Core - 256 bit Bus 901281
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
gary wrote:It is
)
take this with the caveat that it depends a lot on what kind of setup you have.
my 7x and 10x GPU setups use dual fan cards, and perform better (lower temps, faster clock speeds) than they otherwise would with blower model cards. but these systems are not enclosed in a case, but instead on a mining style test bench.
the blower models work well when used in a server style case with high pressure fans, especially if they are right next to each other as the small gap between the cards does not afford good airflow from the suction of the GPU fan alone, it really needs the extra pressure from the chassis fans. i like the density of this kind of server setup, but the hardware to get it done for more than 2-3 GPUs in a single box gets into the realm of proprietary and expensive. spacing more normal hardware out on a mining style setup is much more economical. really the main reason for blowers on "professional" cards is because they are targetted for this kind of server environment with front to back airflow and high pressure fans. also most pro cards have lower clock speeds than their gaming counterparts, making them slightly slower for applications that wont take advantage of the other "pro" features that might exist on such cards, like massive amounts of GPU memory, ECC memory, or front facing power connectors (as opposed to top facing).
i kind of went half way on one system, in which i wanted to get maximum density and ran 7x GPUs plugged directly to the motherboard, but needed to use watercooling to shrink the size to single slot and an external radiator to dump the heat outside the room.
_________________________________________________________________________
After re-reading the "Power
)
After re-reading the "Power Consumption" thread I poked around in the Radeon software I had, found "automatic tuning" with an "under voting" profile.
Now my R5700's are not blowing quite as hot as they were before. And I don't think I am seeing any lost performance (slower processing) either.
That's a win for my power bill and for my A/C trying to keep my house cool during the summer heatwave.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
When the next Gen of Navi
)
When the next Gen of Navi cards comes out should we re-name this thread? Or start up a "Navi 10+ / Big Navi" thread?
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
I think I may have the
)
I think I may have the highest ranked Navi 10 GPU (R5700's) now.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
I need to know what version
)
I need to know what version of Win10 drivers for R5700's (Amd) are successful?
I am getting a steady stream of computational errors as well as invalids on my Win10 4 gpu box.
Tom M
-----Edit-------------
Never mind. It appears I was having a hardware error/malfunctioning gpu. I think my MB is not happy driving 4 gpus :(
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Tom M wrote: I need to know
)
While running 3 gpus on this box ( https://einsteinathome.org/host/12835716 ) I am getting a small number of invalids.
The last one had this message in the output file.
"...
And yesterday I was getting a lot more of the invalids
"...
</stderr_txt>
]]>..."
Any ideas?
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Have you already tried AMD
)
Have you already tried AMD driver version 20.11.1 that was released a couple of days ago ?