the v1.28 app has never worked for Vega on windows and the AMD v1.28 app has been in beta. Something wrong with the windows drivers not providing full support.
mikey, if you disable beta tasks then you’ll get the older app that should work.
I have a whole cache full of SRBase TF tasks right now, they run just fine, and even when I suspend SRBase Einstein won't send me any tasks right now. I have set SRBase to No New Tasks and Einstein TO receive new O3 gpu tasks once they are gone, so probably tomorrow some times I should have some running again.
I DID set the Beta tasks setting to NO so we will see if that fixes the problem, though the card already being CLI 2.0 capable it should work anyway.
Mikey, you might want to reset the project to clear out your client_state.xml file of the offending 1.28 application so that when you connect to the project again you will get the 1.22 app version that works on the Vega cards.
Mikey, you might want to reset the project to clear out your client_state.xml file of the offending 1.28 application so that when you connect to the project again you will get the 1.22 app version that works on the Vega cards.
I DID set the Beta tasks setting to NO so we will see if that fixes the problem, though the card already being CLI 2.0 capable it should work anyway.
"should". but doesn't. and this is only related to the FGRPB1G app. O3MDF tasks shouldn't be a problem.
if you put the same card under Linux with ROCm drivers, it would work. so it's likely a driver problem. the hardware supports OpenCL, but driver implementation seems borked on Vega and older cards. whether it's that they figure no one cares about OpenCL 2.0 on windows with old cards, or it slipped through the cracks and no one has noticed or reported it. Or there could be something wrong with how the Windows app was compiled. either way, the v1.28 AMD Windows app for FGRPB1G is declared beta and does not work with Vega (it does seem to work for RX6000 cards though from other reports).
I DID set the Beta tasks setting to NO so we will see if that fixes the problem, though the card already being CLI 2.0 capable it should work anyway.
"should". but doesn't. and this is only related to the FGRPB1G app. O3MDF tasks shouldn't be a problem.
if you put the same card under Linux with ROCm drivers, it would work. so it's likely a driver problem. the hardware supports OpenCL, but driver implementation seems borked on Vega and older cards. whether it's that they figure no one cares about OpenCL 2.0 on windows with old cards, or it slipped through the cracks and no one has noticed or reported it. Or there could be something wrong with how the Windows app was compiled. either way, the v1.28 AMD Windows app for FGRPB1G is declared beta and does not work with Vega (it does seem to work for RX6000 cards though from other reports).
Thank you that does make sense I guess
Also i can report that the FGRPB!G tasks do work on my Radeon RX6600 gpu
Completed and validated
513
131
3,465
Gamma-ray pulsar binary search #1 on GPUs v1.28 () windows_x86_64
I have found a new Project to use my Vega 56 on...Prime Grid GFN-16 tasks as they literally ZOOM thru them in about 9 to 11 SECONDS for each task and yes they are validating!! For comparison my 3060 is doing them in 1 min 13 seconds each!!
After the end of the Prime Grid TDP challenge this month I will bring it back to either Einstein or SRBase to do tasks, although I'm guessing it may scream thru the Milky Way and Moo tasks as well. I may even try Amicable Numbers as well just to see what happens. The ratio is FP64 = 1/16 FP32 so it's alot better than a 3060Ti which is at FP64 = 1/64 FP32.
I am sure they also have GPU tasks that take longer ;)
I vaguely remember discussions about GPU projects where Radeon gpus were a bunch better at.
I wonder if lhc is one of them?
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 vaguely remember discussions about GPU projects where Radeon gpus were a bunch better at.
few and far between anymore. it seems that for every AMD card that performs well, there's an Nvidia card that will outperform it, if you're willing to pay for it. it wont always be the same card, you have to assess what the workload is doing and what the capabilities of each card is (does it need Integer performance? does it need DP performance? does it need good memory bandwidth? etc).
Einstein was one of them, until the Nvidia apps got fixed/optimized. Asteroids could potentially be a candidate since they rely heavily on DP, but they only have CUDA apps so AMD can't contribute there (same with GPUGRID).
even Milkyway with it's reliance on DP, the leaderboard is littered with older Nvidia cards with good DP performance. it used to be a little cost prohibitive for most people to run the high performing Nvidia DP cards, but the prices of those older Nvidia cards (Titan Vs, P100s) have really come down over the past year or so. a P100 will outperform the Radeon VII on MW, with a lower price for example, but you do require a custom cooling solution if you don't use a server chassis. it only starts to get interesting at the extreme upper echelons of enterprise cards, like the AMD Instinct MI210. but those cards are priced just as far out of reach as the fast Nvidia enterprise cards.
with the driver situation being much better for Nvidia, and the ability to run CUDA-only projects, I find it hard to justify AMD cards anymore unless per GPU cost is your only factor/consideration. personally I try to look at the long term operational costs (electricity) as more important than the one-time purchase price. But everyone's situation is different.
I have found a new Project to use my Vega 56 on...Prime Grid GFN-16 tasks as they literally ZOOM thru them in about 9 to 11 SECONDS for each task and yes they are validating!! For comparison my 3060 is doing them in 1 min 13 seconds each!!
After the end of the Prime Grid TDP challenge this month I will bring it back to either Einstein or SRBase to do tasks, although I'm guessing it may scream thru the Milky Way and Moo tasks as well. I may even try Amicable Numbers as well just to see what happens. The ratio is FP64 = 1/16 FP32 so it's alot better than a 3060Ti which is at FP64 = 1/64 FP32.
mikey, I took a look through your tasks. and you've actually come to the wrong conclusion here.
all of your tasks that run for 10-20 seconds are "proof" tasks. you are only validating that a previously found prime is real. these are very short, and get a tiny amount of credit. your 3060 is doing the proof tasks in like 3 seconds.
if you look at your non-proof tasks, or "main" tasks. you can see that your Vega is actually running about half the speed of your 3060. so the Vega isn't better than the 3060 here either. your Vega is taking about 2.5 minutes on main tasks.
Keith Myers wrote: What does
)
On Boinc startup the event log says:
OpenCL, AMD/ATI gpu 0: RADEON RX Vega(driver version 3444.0(PAL, HSAIL)device version OpenCL! 2.0 AMD=APP (3444.0) 8176MB, 8176 MB available, 11397 GFLOPS peak)
Ian&Steve C. wrote: Oh I
)
I have a whole cache full of SRBase TF tasks right now, they run just fine, and even when I suspend SRBase Einstein won't send me any tasks right now. I have set SRBase to No New Tasks and Einstein TO receive new O3 gpu tasks once they are gone, so probably tomorrow some times I should have some running again.
I DID set the Beta tasks setting to NO so we will see if that fixes the problem, though the card already being CLI 2.0 capable it should work anyway.
Mikey, you might want to
)
Mikey, you might want to reset the project to clear out your client_state.xml file of the offending 1.28 application so that when you connect to the project again you will get the 1.22 app version that works on the Vega cards.
Keith Myers wrote: Mikey,
)
I will do that today thanks
mikey wrote: I DID set the
)
"should". but doesn't. and this is only related to the FGRPB1G app. O3MDF tasks shouldn't be a problem.
if you put the same card under Linux with ROCm drivers, it would work. so it's likely a driver problem. the hardware supports OpenCL, but driver implementation seems borked on Vega and older cards. whether it's that they figure no one cares about OpenCL 2.0 on windows with old cards, or it slipped through the cracks and no one has noticed or reported it. Or there could be something wrong with how the Windows app was compiled. either way, the v1.28 AMD Windows app for FGRPB1G is declared beta and does not work with Vega (it does seem to work for RX6000 cards though from other reports).
_________________________________________________________________________
Ian&Steve C. wrote: mikey
)
Thank you that does make sense I guess
Also i can report that the FGRPB!G tasks do work on my Radeon RX6600 gpu
I have4 found a new Project
)
I have found a new Project to use my Vega 56 on...Prime Grid GFN-16 tasks as they literally ZOOM thru them in about 9 to 11 SECONDS for each task and yes they are validating!! For comparison my 3060 is doing them in 1 min 13 seconds each!!
After the end of the Prime Grid TDP challenge this month I will bring it back to either Einstein or SRBase to do tasks, although I'm guessing it may scream thru the Milky Way and Moo tasks as well. I may even try Amicable Numbers as well just to see what happens. The ratio is FP64 = 1/16 FP32 so it's alot better than a 3060Ti which is at FP64 = 1/64 FP32.
Great!I am sure they also
)
Great!
I am sure they also have GPU tasks that take longer ;)
I vaguely remember discussions about GPU projects where Radeon gpus were a bunch better at.
I wonder if lhc is one of them?
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 vaguely
)
few and far between anymore. it seems that for every AMD card that performs well, there's an Nvidia card that will outperform it, if you're willing to pay for it. it wont always be the same card, you have to assess what the workload is doing and what the capabilities of each card is (does it need Integer performance? does it need DP performance? does it need good memory bandwidth? etc).
Einstein was one of them, until the Nvidia apps got fixed/optimized. Asteroids could potentially be a candidate since they rely heavily on DP, but they only have CUDA apps so AMD can't contribute there (same with GPUGRID).
even Milkyway with it's reliance on DP, the leaderboard is littered with older Nvidia cards with good DP performance. it used to be a little cost prohibitive for most people to run the high performing Nvidia DP cards, but the prices of those older Nvidia cards (Titan Vs, P100s) have really come down over the past year or so. a P100 will outperform the Radeon VII on MW, with a lower price for example, but you do require a custom cooling solution if you don't use a server chassis. it only starts to get interesting at the extreme upper echelons of enterprise cards, like the AMD Instinct MI210. but those cards are priced just as far out of reach as the fast Nvidia enterprise cards.
with the driver situation being much better for Nvidia, and the ability to run CUDA-only projects, I find it hard to justify AMD cards anymore unless per GPU cost is your only factor/consideration. personally I try to look at the long term operational costs (electricity) as more important than the one-time purchase price. But everyone's situation is different.
_________________________________________________________________________
mikey wrote:I have found a
)
mikey, I took a look through your tasks. and you've actually come to the wrong conclusion here.
all of your tasks that run for 10-20 seconds are "proof" tasks. you are only validating that a previously found prime is real. these are very short, and get a tiny amount of credit. your 3060 is doing the proof tasks in like 3 seconds.
if you look at your non-proof tasks, or "main" tasks. you can see that your Vega is actually running about half the speed of your 3060. so the Vega isn't better than the 3060 here either. your Vega is taking about 2.5 minutes on main tasks.
http://www.primegrid.com/results.php?hostid=977404&offset=0&show_names=0&state=2&appid=
_________________________________________________________________________