Previous WU times are variable because there were 2 Vega 64 cards installed in this host. Please note that the 980 sec run time of the previous cards isn't really a fair comparison because I had them running at highly reduced power limits (-45% and 50% if I remember correctly). A more reasonable comparison for Vega 64 vs the VII would be my other host which has 2 Vega 64's running at -25% power targets. It's currently running the same 1048L tasks (3 concurrent WUs) at a 739 sec average compared to about 527 sec average for the VII. So about a 40% improvement in this instance. Hopefully I can improve on that with some tweaking.
Officially AMD has listed only this: PSU Recommendation 750 W
Of course a recent OS version and newest drivers would be a good idea, but I haven't read anything why any motherboard from this decade wouldn't work if it worked with GPUs from a few years back. CPU and RAM speed could have a small effect on the run times, but at least with previous generation cards that difference hasn't been huge.
A more reasonable comparison for Vega 64 vs the VII would be my other host which has 2 Vega 64's running at -25% power targets. It's currently running the same 1048L tasks (3 concurrent WUs) at a 739 sec average compared to about 527 sec average for the VII. So about a 40% improvement in this instance.
Your numbers sound very impressive. I've run only Nvidia for a number of years, and as I take a higher concern for power consumption than most, the Nvidia advantage in power consumption in the Maxwell era seemed pretty compelling.
Among the disappointments of the RTX 20nn cards is that the generational power improvement seems less than customary. (although my 2080 improves quite nicely with undervolting via power limitation). Perhaps AMD has been gaining (overtaking?) on the power efficiency aspect.
Any observed power consumption information you can provide would be very welcome.
... what would be its minimum system requirements?
Uhmmm ... that it POST's when you dust it off and switch it on ??
or perhaps that it has a PCIe slot rather than an AGP slot ?? :-)
Seriously, here is my latest experience which is just a repeat of previous experience. A big online store in Melbourne recently, just happened to have a really good deal (for Australia) on some RX 570s - $AUD 189.00 ea. That's about $US 135.00. I didn't bother checking US prices - irrelevant for the situation here. Seeing as I had paid $AUD 300.00 for the first couple I owned, and later, $AUD 249.00 when they were on special, I couldn't resist $189 now on throwout. So I bought 5 - just to save on the shipping :-).
I still had a couple of long retired clunkers - 2009 vintage. Those now have RX 570s, performing quite well. The only thing I added was a new 450W PSU. I did replace a couple of dodgy looking caps on the (budget Asus with PCIe 1.x) motherboards. The machines still have original CPUs and 2x1G RAM - DDR2 800MHz. I'm not crunching any CPU tasks - just 2xGPU tasks so I figured the low RAM shouldn't matter.
The lowliest machine had an E3200 processor - that's a 2.4GHz Celeron dual core. Here is the details page. It's only been running a couple of days. I tested some tasks singly - found ~660 secs crunch time - before switching to x2. This gave 2 tasks in ~1180 secs - 590 secs per task.
By way of comparison, I have a brand new build (around 7 months ago) using a G4560 Kaby Lake processor - dual core, 4 threads, 3.5GHz, 2x4GB DDR4 2400MHz RAM in an Asrock B250M PRO4 board. This machine also, is not running any CPU tasks. It was running a different GPU but I decided to swap to one of the new 570s so I could really see what difference the board/CPU/RAM/PCIe generation actually made. For this one, the details page is here. I was quite interested to see that the x2 crunch times are also pretty much ~1180 secs - 590 secs per task. The new GPU was added on 11th Feb so results before then would be from a Pitcairn series R7 370 GPU.
These two machines now have exactly the same OS and kernel, the same amdgpu driver and OpenCL libs, the same BOINC version (my home built 7.14.2) driving the exact same model RX 570 GPU. The identical crunch times clearly indicate that the motherboard, RAM and CPU (or even PCIe version) don't have much effect on crunch times.
I'm not saying your experience with a Radeon VII would also not be affected by those same things. However it seems likely that you would get similar performance without having to shell out for any more expensive supporting hardware. The thing not to skimp on would be a decent PSU :-).
Right now I'm crunching at a little better than 1.7M /day and pulling about 450W at the wall with a Gold rated 850W PSU. The VII is lightly under-volted with power limit % of -20 (which is the lowest Wattman currently allows on the VII). Hopefully as the drivers get improved it will be possible to drop the power limit further. I've not yet played with reducing the max clock speed. I know this can be improved upon, but am just going to let it run until I have more time to play around.
I used to run Nvidia cards, but now that Einstein is my primary focus it's a no brainier to use AMD. For this application their cards much more powerful for the $ and can also be more efficient if properly tuned.
Hey, congrats for being the first here with the VII! :)
The machine looks nice, to get to 1,7 mil with one card is awesome. Still, have you thought about adding one more, to crunch with 2 VIIs? The CPU seems capable. And even the PSU can hold (300Wx2), even if a bigger one would be recommended.
The identical crunch times clearly indicate that the motherboard, RAM and CPU (or even PCIe version) don't have much effect on crunch times.
Great info on hardware requirements Gary. I've also performed some testing with running Vega 64's in PCIe 3.0 x16, x8, x4 and x1. I couldn't detect any noticeable performance loss using x8 or x4 slots vs. the x16. Even an x1 slot worked reasonably well; it was only a couple % slower, and that's with a high end card.
Hey, congrats for being the first here with the VII! :)
Thanks, but I wasn't the first, just the first to post about it, there are a few others lurking out there!
John wrote:
The machine looks nice, to get to 1,7 mil with one card is awesome. Still, have you thought about adding one more, to crunch with 2 VIIs? The CPU seems capable. And even the PSU can hold (300Wx2), even if a bigger one would be recommended.
Of course I dreamed about how great it would be to have have another VII, especially since I've previously had 3 Vega's in this same machine. There's a few problems though!
Short term, the VII is out of stock everywhere right now except marked up on eBay.
The tri-fan cooling design isn't great for stuffing multiple cards into one machine and it's the only option right now.
FYI, I have a developer friend that has put together a Linux utility for controlling AMD/ATI cards. It can control clocks, power levels and p-states. Seeing requests here frequently for the ability to undervolt AMD cards for power usage savings, I thought some might find Rick's utility helpful.
Previous WU times are
)
Previous WU times are variable because there were 2 Vega 64 cards installed in this host. Please note that the 980 sec run time of the previous cards isn't really a fair comparison because I had them running at highly reduced power limits (-45% and 50% if I remember correctly). A more reasonable comparison for Vega 64 vs the VII would be my other host which has 2 Vega 64's running at -25% power targets. It's currently running the same 1048L tasks (3 concurrent WUs) at a 739 sec average compared to about 527 sec average for the VII. So about a 40% improvement in this instance. Hopefully I can improve on that with some tweaking.
Vega 64 host for comparison:
https://einsteinathome.org/host/12667735
If an envious soul were to
)
If an envious soul were to buy or build a host to run a Vega VII, what would be its minimum system requirements?
Ideas are not fixed, nor should they be; we live in model-dependent reality.
Officially AMD has listed
)
mesman21 wrote:A more
)
Your numbers sound very impressive. I've run only Nvidia for a number of years, and as I take a higher concern for power consumption than most, the Nvidia advantage in power consumption in the Maxwell era seemed pretty compelling.
Among the disappointments of the RTX 20nn cards is that the generational power improvement seems less than customary. (although my 2080 improves quite nicely with undervolting via power limitation). Perhaps AMD has been gaining (overtaking?) on the power efficiency aspect.
Any observed power consumption information you can provide would be very welcome.
cecht wrote:... what would be
)
Uhmmm ... that it POST's when you dust it off and switch it on ??
or perhaps that it has a PCIe slot rather than an AGP slot ?? :-)
Seriously, here is my latest experience which is just a repeat of previous experience. A big online store in Melbourne recently, just happened to have a really good deal (for Australia) on some RX 570s - $AUD 189.00 ea. That's about $US 135.00. I didn't bother checking US prices - irrelevant for the situation here. Seeing as I had paid $AUD 300.00 for the first couple I owned, and later, $AUD 249.00 when they were on special, I couldn't resist $189 now on throwout. So I bought 5 - just to save on the shipping :-).
I still had a couple of long retired clunkers - 2009 vintage. Those now have RX 570s, performing quite well. The only thing I added was a new 450W PSU. I did replace a couple of dodgy looking caps on the (budget Asus with PCIe 1.x) motherboards. The machines still have original CPUs and 2x1G RAM - DDR2 800MHz. I'm not crunching any CPU tasks - just 2xGPU tasks so I figured the low RAM shouldn't matter.
The lowliest machine had an E3200 processor - that's a 2.4GHz Celeron dual core. Here is the details page. It's only been running a couple of days. I tested some tasks singly - found ~660 secs crunch time - before switching to x2. This gave 2 tasks in ~1180 secs - 590 secs per task.
By way of comparison, I have a brand new build (around 7 months ago) using a G4560 Kaby Lake processor - dual core, 4 threads, 3.5GHz, 2x4GB DDR4 2400MHz RAM in an Asrock B250M PRO4 board. This machine also, is not running any CPU tasks. It was running a different GPU but I decided to swap to one of the new 570s so I could really see what difference the board/CPU/RAM/PCIe generation actually made. For this one, the details page is here. I was quite interested to see that the x2 crunch times are also pretty much ~1180 secs - 590 secs per task. The new GPU was added on 11th Feb so results before then would be from a Pitcairn series R7 370 GPU.
These two machines now have exactly the same OS and kernel, the same amdgpu driver and OpenCL libs, the same BOINC version (my home built 7.14.2) driving the exact same model RX 570 GPU. The identical crunch times clearly indicate that the motherboard, RAM and CPU (or even PCIe version) don't have much effect on crunch times.
I'm not saying your experience with a Radeon VII would also not be affected by those same things. However it seems likely that you would get similar performance without having to shell out for any more expensive supporting hardware. The thing not to skimp on would be a decent PSU :-).
Cheers,
Gary.
Right now I'm crunching at a
)
Right now I'm crunching at a little better than 1.7M /day and pulling about 450W at the wall with a Gold rated 850W PSU. The VII is lightly under-volted with power limit % of -20 (which is the lowest Wattman currently allows on the VII). Hopefully as the drivers get improved it will be possible to drop the power limit further. I've not yet played with reducing the max clock speed. I know this can be improved upon, but am just going to let it run until I have more time to play around.
I used to run Nvidia cards, but now that Einstein is my primary focus it's a no brainier to use AMD. For this application their cards much more powerful for the $ and can also be more efficient if properly tuned.
Hey, congrats for being the
)
Hey, congrats for being the first here with the VII! :)
The machine looks nice, to get to 1,7 mil with one card is awesome. Still, have you thought about adding one more, to crunch with 2 VIIs? The CPU seems capable. And even the PSU can hold (300Wx2), even if a bigger one would be recommended.
Gary Roberts wrote:The
)
Great info on hardware requirements Gary. I've also performed some testing with running Vega 64's in PCIe 3.0 x16, x8, x4 and x1. I couldn't detect any noticeable performance loss using x8 or x4 slots vs. the x16. Even an x1 slot worked reasonably well; it was only a couple % slower, and that's with a high end card.
John wrote:Hey, congrats for
)
Thanks, but I wasn't the first, just the first to post about it, there are a few others lurking out there!
Of course I dreamed about how great it would be to have have another VII, especially since I've previously had 3 Vega's in this same machine. There's a few problems though!
FYI, I have a developer
)
FYI, I have a developer friend that has put together a Linux utility for controlling AMD/ATI cards. It can control clocks, power levels and p-states. Seeing requests here frequently for the ability to undervolt AMD cards for power usage savings, I thought some might find Rick's utility helpful.
https://github.com/Ricks-Lab/amdgpu-utils
He would like feedback and requests for new features.