System also has a Radeon VII and a pair of Vega 64's, and they are crunching away on this driver without issue.
That's unfortunate about the invalids. Assuming that will get worked out eventually, how do task completion times compare among the 5700XT, RadeonVI and Vega64?
Don't believe so. That is only the firmware needed to detect and support the cards. The problem with the cards are the OpenCL components of the drivers which are provided by AMD and the Khronos Group. As far as I know they still haven't figured out how to make compute work in the drivers for these cards.
is there an update about this topic? conssidering to buy a RX 5700 XT but I also want to be able to run einstein.
I just ran a few tasks on my 5700XT and one of them already came back invalid, so I would say no. There will either need to be a new app from Einstein, or updated drivers from AMD, before Einstein will run on Navi.
After several weeks of gathering dust I fired up my RX 5700 XT (Win 10 Pro) machine for a few hours this last weekend, upgraded to driver 19.10.2 and set about cunching some GRP work at x1. Very quickly I was seeing completed tasks being declared invalid so switched to GW work (again at x1) in the hope that Bernd's magic on that app would pay dividends... 16 - 18 hours later I aborted what was left in the work cache and shut the machine down again having built a raft of pending validation results.
Having just checked to see if any of my work testing was worthwhile I have discovered that all of my sample GRP tasks were declared invalid (no surprise there) but so far I have 18 validated GW tasks and that is by far and away the most succesful this card has ever been! I still have something like 70 odd (mostly GW) tasks pending validation so will delay restarting the machine for a day or two to see whether those pendings become valid too before I return this machine to full time running on GW.
G all of my sample GRP tasks were declared invalid
<snip>
18 validated GW tasks
That is very encouraging. It seems to me to more strongly support the notion that some Navi numeric results are very slightly different, rather than the alternate hypothesis widely circulated that AMD just has not got around to getting openCL right for this architecture.
Awkwardly, that leaves us not much grounds to guess whether results it will produce on future Einstein applications will be judged sufficiently similar to results from other platforms to validate in quorums of not the same platform.
After several weeks of gathering dust I fired up my RX 5700 XT (Win 10 Pro) machine for a few hours this last weekend, upgraded to driver 19.10.2 and set about cunching some GRP work at x1. Very quickly I was seeing completed tasks being declared invalid so switched to GW work (again at x1) in the hope that Bernd's magic on that app would pay dividends... 16 - 18 hours later I aborted what was left in the work cache and shut the machine down again having built a raft of pending validation results.
Having just checked to see if any of my work testing was worthwhile I have discovered that all of my sample GRP tasks were declared invalid (no surprise there) but so far I have 18 validated GW tasks and that is by far and away the most succesful this card has ever been! I still have something like 70 odd (mostly GW) tasks pending validation so will delay restarting the machine for a day or two to see whether those pendings become valid too before I return this machine to full time running on GW.
Gav.
I'll fire up some GW tasks on my 5700XT in Linux and see how it goes. I don't think I have tried to run any of those recently.
Not sure if it's the Linux driver or the GW app. I'll update once the rest of my pending tasks go invalid or validate.
In view of your continued lack of success I'm almost ashamed to report that so far none of my previous run of GW tasks have failed to validate against their CPU counterparts so it's good news for me!
Encouraged by this I restrated the machine a day or so ago and I'm now running the new O2MDF flavour of GW work being validated against NVidia GPU's and although it is early days things are looking good with several tasks already validated and no invalids or errors for the current run. I am now tempting fate by adjusting preferences to allow tasks to run x2 in the hope of increasing the very low usage of the 5700XT... I'll be back as that man once said with further reports.
So, I just got my 5700XT 2 days ago, and the results are pretty concerning to me.
3 tasks were marked as valid and 3 tasks were marked as invalid. However, the valid tasks were marked as valid only when compared to a result from a system with another 5700XT!
cecht wrote:tictoc
)
Runtimes running one task at a time per GPU:
Radeon VII @ 1940core|1175mem - 190 seconds
Vega 64 @ 1650core|1025mem- 280 seconds
5700XT @ 1970core|1750mem - 330 seconds
Does this news from Phoronix
)
Does this news from Phoronix affect RX5700s crunching for E@H?
Ideas are not fixed, nor should they be; we live in model-dependent reality.
Don't believe so. That is
)
Don't believe so. That is only the firmware needed to detect and support the cards. The problem with the cards are the OpenCL components of the drivers which are provided by AMD and the Khronos Group. As far as I know they still haven't figured out how to make compute work in the drivers for these cards.
hi, is there an update
)
hi,
is there an update about this topic? conssidering to buy a RX 5700 XT but I also want to be able to run einstein.
Timo wrote:hi,is there an
)
I just ran a few tasks on my 5700XT and one of them already came back invalid, so I would say no. There will either need to be a new app from Einstein, or updated drivers from AMD, before Einstein will run on Navi.
Glimmer of hope
)
Glimmer of hope perhaps.
After several weeks of gathering dust I fired up my RX 5700 XT (Win 10 Pro) machine for a few hours this last weekend, upgraded to driver 19.10.2 and set about cunching some GRP work at x1. Very quickly I was seeing completed tasks being declared invalid so switched to GW work (again at x1) in the hope that Bernd's magic on that app would pay dividends... 16 - 18 hours later I aborted what was left in the work cache and shut the machine down again having built a raft of pending validation results.
Having just checked to see if any of my work testing was worthwhile I have discovered that all of my sample GRP tasks were declared invalid (no surprise there) but so far I have 18 validated GW tasks and that is by far and away the most succesful this card has ever been! I still have something like 70 odd (mostly GW) tasks pending validation so will delay restarting the machine for a day or two to see whether those pendings become valid too before I return this machine to full time running on GW.
Gav.
Gavin wrote:G all of my
)
That is very encouraging. It seems to me to more strongly support the notion that some Navi numeric results are very slightly different, rather than the alternate hypothesis widely circulated that AMD just has not got around to getting openCL right for this architecture.
Awkwardly, that leaves us not much grounds to guess whether results it will produce on future Einstein applications will be judged sufficiently similar to results from other platforms to validate in quorums of not the same platform.
Thanks for sharing your observations.
Gavin wrote:Glimmer of hope
)
I'll fire up some GW tasks on my 5700XT in Linux and see how it goes. I don't think I have tried to run any of those recently.
**Edit**
No luck so far for me on my 5700XT. Everything is coming back invalid so far. https://einsteinathome.org/host/12492619/tasks/5/0
Not sure if it's the Linux driver or the GW app. I'll update once the rest of my pending tasks go invalid or validate.
tictoc wrote: No luck so far
)
In view of your continued lack of success I'm almost ashamed to report that so far none of my previous run of GW tasks have failed to validate against their CPU counterparts so it's good news for me!
Encouraged by this I restrated the machine a day or so ago and I'm now running the new O2MDF flavour of GW work being validated against NVidia GPU's and although it is early days things are looking good with several tasks already validated and no invalids or errors for the current run. I am now tempting fate by adjusting preferences to allow tasks to run x2 in the hope of increasing the very low usage of the 5700XT... I'll be back as that man once said with further reports.
Gav.
So, I just got my 5700XT 2
)
So, I just got my 5700XT 2 days ago, and the results are pretty concerning to me.
3 tasks were marked as valid and 3 tasks were marked as invalid. However, the valid tasks were marked as valid only when compared to a result from a system with another 5700XT!
This workunit: https://einsteinathome.org/workunit/426141023 and this workunit: https://einsteinathome.org/workunit/426059969 are such examples.
I am pausing my computations as I'm worried that my results are actually invalid. I don't want my PC to produce results that are bad for the project.