I would stop using RAC as an indicator for performance. It takes a long time to stabilize and small changes can affect RAC less than the normal verification variation of tasks. Just go off task run time. If one task runs for 500s, 2x run for 990s and 3x are 1485s each then there is an improvement.
Generally I agree, with the reservation that one needs also to look at the invalid rate. In recent years my invalid rate on GRP work has not varied much outside the band of 0.5% to 1.5%, and usually not shown obvious systematic tendencies by model of card or degree of multiplicity.
However, when I boosted my VII from 2X to 3X, the already somewhat above normal invalid rate jumped sufficiently to more than consume the productivity improvement. I'm not claiming that is a general characteristic of the VII cards--I think mine was a bit substandard in one or more ways.
But it is an illustration of why one wants to take a look. Mostly it is not informative, but the tiny effort to look at the relative size of the summary numbers at the top of the task page is worth it, I think.
Ok, I now have a "full rack" (eg. 6 XFX Rx 570/580) cards.
My goal is to expand upto 9 Rx 570 8gb cards. It turns out XFX cards are more expensive (and faster). Since I expect to either run 2 GR or 2 GW tasks they pretty much have to be 8gb cards.
There are some Rx 470/480 8GB cards out there that are cheaper.
Sigh.
Any advice?
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 have taken it on faith (and measurement) for some time that my RX 570 was faster, and more importantly more efficient than my Nvidia cards, on both GW and FGRP. But times change.
For FGRP,
I just tried a GTX 1060 under Ubuntu 20.04.1 (460 driver), and am seeing times of under 14 minutes, at a power of about 75 watts (nvidia-smi -l).
I don't have a recent measurement for my RX 570, but my previous ones show it slightly faster though at more power, so the Nvidias look to be more efficient.
My Rx 570/580 cards seem to be running 20+ minutes per GW gpu task. (single threaded)
And around 7.5 to 10.5 minutes for Gamma Ray in earlier iterations. (single threaded)
The most important issue used to be the Rx cards will run more threads at the same time for the GR tasks than the NVIDIA cards will.
Before I bailed on most of my NVIDIA cards I wasn't gaining anything by running 2 threads. But the Rx cards were gaining by running multiple tasks.
GW however requires a lot more CPU support and apparently more video memory than GR does.
For the Rx 570/580 you may be able to run 2 GR tasks. Or perhaps higher.
But for the GW tasks 2 maybe the upper limit on Rx 570/580 cards.
GR and GW gpu tasks do NOT mix very well. Run one or the other. But don't enable both at the same time on a profile. Don't remember if just switching between them is less of a problem. Think it was less of a problem.
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 don't have a recent measurement for my RX 570, but my previous ones show it slightly faster ...
You need to be aware that crunch times have changed for everybody with a new type of task that started being used a couple of weeks ago. All tasks with names that start LATeah3001 (and above) run a *lot* faster (so far) than the previous types.
As an example, one of my RX 570s used to take around 10 mins per task with the former type. It now takes just 7 mins. An R7 370 used to take over 20 mins. It now takes less than 13 mins.
I don't run any nvidia GPUs but I imagine those who do are seeing similar output gains. The new tasks obviously have a significantly lower work content. Enjoy them while they last :-).
I would stop using RAC as an
)
I would stop using RAC as an indicator for performance. It takes a long time to stabilize and small changes can affect RAC less than the normal verification variation of tasks. Just go off task run time. If one task runs for 500s, 2x run for 990s and 3x are 1485s each then there is an improvement.
mmonnin wrote:Just go off
)
Generally I agree, with the reservation that one needs also to look at the invalid rate. In recent years my invalid rate on GRP work has not varied much outside the band of 0.5% to 1.5%, and usually not shown obvious systematic tendencies by model of card or degree of multiplicity.
However, when I boosted my VII from 2X to 3X, the already somewhat above normal invalid rate jumped sufficiently to more than consume the productivity improvement. I'm not claiming that is a general characteristic of the VII cards--I think mine was a bit substandard in one or more ways.
But it is an illustration of why one wants to take a look. Mostly it is not informative, but the tiny effort to look at the relative size of the summary numbers at the top of the task page is worth it, I think.
Ok, I now have a "full rack"
)
Ok, I now have a "full rack" (eg. 6 XFX Rx 570/580) cards.
My goal is to expand upto 9 Rx 570 8gb cards. It turns out XFX cards are more expensive (and faster). Since I expect to either run 2 GR or 2 GW tasks they pretty much have to be 8gb cards.
There are some Rx 470/480 8GB cards out there that are cheaper.
Sigh.
Any advice?
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
From the TechPowerUp website
)
From the TechPowerUp website database, the RX 480 and RX570 look to have almost identical specs. Just a slight FP64 advantage to the RX480.
https://www.techpowerup.com/gpu-specs/radeon-rx-480.c2848
https://www.techpowerup.com/gpu-specs/radeon-rx-570.c2939
580 = Overclocked 480 570
)
580 = Overclocked 480
570 = Overclocked 470
470 is a 480 chip with part disabled.
https://www.anandtech.com/show/11278/amd-radeon-rx-580-rx-570-review
mmonnin wrote: 580 =
)
Thank you MMONNIN and Keith!
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 have taken it on faith (and
)
I have taken it on faith (and measurement) for some time that my RX 570 was faster, and more importantly more efficient than my Nvidia cards, on both GW and FGRP. But times change.
For FGRP,
I just tried a GTX 1060 under Ubuntu 20.04.1 (460 driver), and am seeing times of under 14 minutes, at a power of about 75 watts (nvidia-smi -l).
https://einsteinathome.org/host/12790643/tasks/0/0
And a GTX 1650 Super under Win10 (457.51 driver) is running slightly faster at a power of 72 watts (GPU-Z board power).
https://einsteinathome.org/host/12871756/tasks/0/0
I don't have a recent measurement for my RX 570, but my previous ones show it slightly faster though at more power, so the Nvidias look to be more efficient.
I have not tried GW yet.
My Rx 570/580 cards seem to
)
My Rx 570/580 cards seem to be running 20+ minutes per GW gpu task. (single threaded)
And around 7.5 to 10.5 minutes for Gamma Ray in earlier iterations. (single threaded)
The most important issue used to be the Rx cards will run more threads at the same time for the GR tasks than the NVIDIA cards will.
Before I bailed on most of my NVIDIA cards I wasn't gaining anything by running 2 threads. But the Rx cards were gaining by running multiple tasks.
GW however requires a lot more CPU support and apparently more video memory than GR does.
For the Rx 570/580 you may be able to run 2 GR tasks. Or perhaps higher.
But for the GW tasks 2 maybe the upper limit on Rx 570/580 cards.
GR and GW gpu tasks do NOT mix very well. Run one or the other. But don't enable both at the same time on a profile. Don't remember if just switching between them is less of a problem. Think it was less of a problem.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Jim1348 wrote:... I don't
)
You need to be aware that crunch times have changed for everybody with a new type of task that started being used a couple of weeks ago. All tasks with names that start LATeah3001 (and above) run a *lot* faster (so far) than the previous types.
As an example, one of my RX 570s used to take around 10 mins per task with the former type. It now takes just 7 mins. An R7 370 used to take over 20 mins. It now takes less than 13 mins.
I don't run any nvidia GPUs but I imagine those who do are seeing similar output gains. The new tasks obviously have a significantly lower work content. Enjoy them while they last :-).
Cheers,
Gary.
Is anyone else trolling for
)
Is anyone else trolling for this kind of Rx 400/500 series card?
Has anyone else found a supply of 8GB cards that are exceptionally low cost?
I am assuming the 4GB cards will not be able to run two GW tasks which is where I am shooting for with 6-9 GPUs (have 6, need 3).
I really don't want to spend ~$300 on a card that has half the production of the Rx 5700 cards I got (for ~$300 back before the gpu price inflation).
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)