Testing undervolting on the 5700 XT from 1.2 to 1.1V. Power reported reduced by ~20 W. So far so good.
I've gone the other way with my testing. Nearly 48 hours ago I switched to running Gamma-ray tasks x2, tweaked the fan curve, raised the memory speed slightly and applied +10 to the power limit...
GPU clock hovers around 1930Mhz, Temps are stable at 63C (the machine is in a cold unheated room), memory speed is 1800Mhz and board power is mostly 160W with a few highs and lows.
In this trim, performance comparisions can be drawn from my undervolted Vega64 cards and it seems the cards are equally matched in terms of cunch times and throughput with the 5700XT winning on the power consumption front by 60 or more Watts without undervolting.
In the few days since upgrading to this new driver I have completed nearly 1000 Gamma-ray jobs and have only picked up 5 invalids, they are from the period of running x1 concurrency and a simple tweak to the GPU fan curve. More testing is needed to find the perfect balance between productivity, reliability and energy consumption but so far things are on the up for Navi here at Einstein. I hope we hear news from other Navi owners and their experiences to add to our fledgling knowledge of these cards behaviour :-)
Nearly 48 hours ago I switched to running Gamma-ray tasks x2, tweaked the fan curve, raised the memory speed slightly and applied +10 to the power limit...Gav.
Wow.
You were logging a pretty consistent task elapsed time near 338 seconds at 1X before the changes, and quickly settled to about 570 seconds at 2X.
Do you have an estimate of how much of that improvement came from higher clock rates, and how much was just the 2X bonus? If I take that all as 2X bonus it would be an 18.5% productivity improvement, which is more than I am used to seeing in recent years on Einstein GRP work.
For comparison, my RX 570 cards running Einstein GRP under Windows 10 generate 2X elapsed times of about 1200 seconds on one machine, and 1215 on the other, both with -35% power limitation limit in place (but that limit is not costing much in productivity). So your most recent results show a nice fat productivity advantage over my 570 of a bit over a factor of 2.1. Even scaling back a little to comparable power settings, I'll guess there is be a real advantage near a factor of 2.
Of course, my very recent XFX RX 570 card purchases cost me $140, while a current price from the same source for the same brand of 5700 is $320, and $380 for the 5700 XT.
But considering both the capital cost and the power consumption of the host, getting a lot more Einstein GRP work out of a host PCI slot allows one to break even at a card cost/daily credit and card power/credit rate appreciably below break even.
I wish I knew the advantage for Einstein GRP of the 5700 XT over the 5700. Some published material leads me to suspect there may be a material power savings from the non-XT model.
You were logging a pretty consistent task elapsed time near 338 seconds at 1X before the changes, and quickly settled to about 570 seconds at 2X.
Do you have an estimate of how much of that improvement came from higher clock rates, and how much was just the 2X bonus? If I take that all as 2X bonus it would be an 18.5% productivity improvement, which is more than I am used to seeing in recent years on Einstein GRP work.
Runtimes are very tight with only a second or two of variation, which is good. I'm guessing the variation is due to cpu support fluctuating slightly between tasks. It's probably also worth noting that the 5700 XT is living in a pci-e2 slot on a ageing low end motherboard with a equally old middle of the road quad core cpu.
I've rechecked my notes as I made a mistake with my earlier post. My fan curve adjustment and memory overclock were applied for the tasks run at x1 and were carried over to my x2 testing when I applied the +10 power limit because of a couple of lock ups. Since then all has been stable.
I would suggest the productivity improvement is mostly due to running tasks x2. I will probably leave things as they are for a few days then maybe after the weekend try running x3.
Testing undervolting on the 5700 XT from 1.2 to 1.1V. Power reported reduced by ~20 W. So far so good.
You had some big batches of compute errors returned January 20 and 21. Did you find the undervolt limit?
Don't know what exactly happened, but most probably some driver crash. I noticed this much later only, hence the large amount of errors. The system was working, but artifacts on screen. I did a driver update and during it the screen went back to normal, but GPU tasks were still failing until I finally rebooted. Might not be due to undervolting at all. Running again @ 1.1V, no issues since then. Will keep an eye on it.
BTW, this system is configured to process GPU work only during 5pm-8am on weekdays, so it doesn't disturb me during work. So don't rely on the RAC, it doesn't reflect the real potential.
Only half a year after launch, I am able to use my 5700 XT on Einstein@home. 824s elapsed time on 3x. 275s per WU.
While for Radeon VII, 320s elapsed time on 2x. 160s per WU. Considering 5700 XT is priced at $400 and Radeon VII at $550. Radeon VII is still the king.
Inspired by Mumak and Gavin's reports of success in running Navi 10 5700 XT cards on the latest AMD drivers, and motivated by some serious problems on my primary personal computing and Einstein crunching box, I purchased and today installed an RX 5700 card (NOT XT).
The system in question has had a varied recent history. I ran a Radeon VII on it for many months last year, but reached a point of 100% Einstein invalid results in December and pulled the card. I installed first one RX 570 on it, then the day before Christmas added a second RX 570. I got lots of valid results from the double 570 configuration, but also a huge system stability headache (more blue screens in one day than my whole flotilla normally produces in a year). I currently blame "something" about the second 570, as when I pulled it the system instantly became stable and has remained so.
Initial impressions: This thing is BIG. It is the thickest card I've seen (termed a 2.5 slot card) and much longer and heavier than my XFX RX 570 cards.
The card has a tiny switch resembling the famous "mining mode BIOS" switch on the XFX RX 570 cards I've used. A little Internet searching suggested that also on this 5700 card XFX might have configured the forward (toward the I/O panel) switch position for a power-efficient mode of operation. I took a chance and moved that switch to that position on first install. Thus all results I post here may be non-representative of typical 5700 card behavior.
My very first impression was astonishment at how very little more power the card was burning when running Einstein than had my RX 570.
My second impression is that nevertheless there is a very considerable performance gain over the 570. Thus I am seeing good power efficiency gain at the card level, and a very good system level power efficiency gain--especially if I am later able to add back in my healthy 570, and perhaps, a little later, make this a dual 5700 system.
I ran a single WU at 1X. As I chose a _2 task, I got instant validation, which was comforting. The elapsed time running 1X was 6:12. I started running 2X partway through the second task, and am currently running steady 2X with elapsed times running about 10:47. That is a nominal daily credit rate of 925,000. Of course, in real life, the productivity will be less, as I must expect at least 1% loss to validation, and as this system is my daily driver, there will be also loss to task interference, down time, etc.
As I have complained about my VII on fan noise grounds, I should mention that this one is so far running the GPU fan at 33%, or 1153 rpm as reported by GPU-Z. It is not very noisy, although there is a noticeable surge at the end of a task. The character of the fan sound is quite pleasing (not a whine, beat tones, ... --think Noctua-like).
It is not only early days, but early minutes (and AMD has only had a driver release which would give Einstein GRP validations very recently).
I consider all of these results very preliminary, subject to adverse revision.
Current operating conditions (as reported by HWiNFO, and confirmed with GPU-Z)
AMD driver version: 20.1.2 released on January 20, 2020
GPU clock 1736 Max, 1707 Average
Memory clock 1750 unchanging
GPU core voltage 0.956 nearly always
GPU Temperature 76 C average
GPU Hot Spot Temperature 88 average
GPU fan 1810 peak, 1167 average
The approximate PC power draw at the wall socket is 225W--I'll take a good long-term average measurement later. For comparison, my (last night) good long-term average measure running a single RX 570 card at -35% power limitation in the same box was 188W!! At that condition the indicated daily productivity with the RX 570 was only 490,000.
If this holds together, I am very pleased. As the box has a Seasonic 850W supply and lots of fans, I'm also pretty optimistic about getting another card running in it (at least a 570, quite possibly another 5700).
I suspect the very low operating voltage is a big factor in the good power efficiency I am seeing. I don't know how much of that is because it is a 5700, and not a 5700 XT, how much because of the BIOS switch in "mining" position, or something else.
I started running 2X partway through the second task, and am currently running steady 2X with elapsed times running about 10:47. That is a nominal daily credit rate of 925,000.
Congrats on your new card's success! I'm liking everything you've posted about it so far. I've been seriously considering getting that same card. Let us know how 3X works. And 4X!
Has anybody had success running gravity wave work on the 5700-series?
Ideas are not fixed, nor should they be; we live in model-dependent reality.
Testing undervolting on the
)
Testing undervolting on the 5700 XT from 1.2 to 1.1V. Power reported reduced by ~20 W. So far so good.
-----
Mumak wrote:Testing
)
I've gone the other way with my testing. Nearly 48 hours ago I switched to running Gamma-ray tasks x2, tweaked the fan curve, raised the memory speed slightly and applied +10 to the power limit...
GPU clock hovers around 1930Mhz, Temps are stable at 63C (the machine is in a cold unheated room), memory speed is 1800Mhz and board power is mostly 160W with a few highs and lows.
In this trim, performance comparisions can be drawn from my undervolted Vega64 cards and it seems the cards are equally matched in terms of cunch times and throughput with the 5700XT winning on the power consumption front by 60 or more Watts without undervolting.
In the few days since upgrading to this new driver I have completed nearly 1000 Gamma-ray jobs and have only picked up 5 invalids, they are from the period of running x1 concurrency and a simple tweak to the GPU fan curve. More testing is needed to find the perfect balance between productivity, reliability and energy consumption but so far things are on the up for Navi here at Einstein. I hope we hear news from other Navi owners and their experiences to add to our fledgling knowledge of these cards behaviour :-)
Gav.
Gavin wrote:Nearly 48 hours
)
Wow.
You were logging a pretty consistent task elapsed time near 338 seconds at 1X before the changes, and quickly settled to about 570 seconds at 2X.
Do you have an estimate of how much of that improvement came from higher clock rates, and how much was just the 2X bonus? If I take that all as 2X bonus it would be an 18.5% productivity improvement, which is more than I am used to seeing in recent years on Einstein GRP work.
For comparison, my RX 570 cards running Einstein GRP under Windows 10 generate 2X elapsed times of about 1200 seconds on one machine, and 1215 on the other, both with -35% power limitation limit in place (but that limit is not costing much in productivity). So your most recent results show a nice fat productivity advantage over my 570 of a bit over a factor of 2.1. Even scaling back a little to comparable power settings, I'll guess there is be a real advantage near a factor of 2.
Of course, my very recent XFX RX 570 card purchases cost me $140, while a current price from the same source for the same brand of 5700 is $320, and $380 for the 5700 XT.
But considering both the capital cost and the power consumption of the host, getting a lot more Einstein GRP work out of a host PCI slot allows one to break even at a card cost/daily credit and card power/credit rate appreciably below break even.
I wish I knew the advantage for Einstein GRP of the 5700 XT over the 5700. Some published material leads me to suspect there may be a material power savings from the non-XT model.
Thanks for sharing your results.
archae86 wrote: Wow. You
)
Runtimes are very tight with only a second or two of variation, which is good. I'm guessing the variation is due to cpu support fluctuating slightly between tasks. It's probably also worth noting that the 5700 XT is living in a pci-e2 slot on a ageing low end motherboard with a equally old middle of the road quad core cpu.
I've rechecked my notes as I made a mistake with my earlier post. My fan curve adjustment and memory overclock were applied for the tasks run at x1 and were carried over to my x2 testing when I applied the +10 power limit because of a couple of lock ups. Since then all has been stable.
I would suggest the productivity improvement is mostly due to running tasks x2. I will probably leave things as they are for a few days then maybe after the weekend try running x3.
Gav.
Mumak wrote:Testing
)
You had some big batches of compute errors returned January 20 and 21. Did you find the undervolt limit?
archae86 wrote:Mumak
)
Don't know what exactly happened, but most probably some driver crash. I noticed this much later only, hence the large amount of errors. The system was working, but artifacts on screen. I did a driver update and during it the screen went back to normal, but GPU tasks were still failing until I finally rebooted. Might not be due to undervolting at all. Running again @ 1.1V, no issues since then. Will keep an eye on it.
BTW, this system is configured to process GPU work only during 5pm-8am on weekdays, so it doesn't disturb me during work. So don't rely on the RAC, it doesn't reflect the real potential.
-----
Only half a year after
)
Only half a year after launch, I am able to use my 5700 XT on Einstein@home. 824s elapsed time on 3x. 275s per WU.
While for Radeon VII, 320s elapsed time on 2x. 160s per WU. Considering 5700 XT is priced at $400 and Radeon VII at $550. Radeon VII is still the king.
I don't see a comparison
)
I don't see a comparison between Windows and Linux. I am interested in Fermi Gamma Ray, on any AMD GPU.
Inspired by Mumak and Gavin's
)
Inspired by Mumak and Gavin's reports of success in running Navi 10 5700 XT cards on the latest AMD drivers, and motivated by some serious problems on my primary personal computing and Einstein crunching box, I purchased and today installed an RX 5700 card (NOT XT).
The system in question has had a varied recent history. I ran a Radeon VII on it for many months last year, but reached a point of 100% Einstein invalid results in December and pulled the card. I installed first one RX 570 on it, then the day before Christmas added a second RX 570. I got lots of valid results from the double 570 configuration, but also a huge system stability headache (more blue screens in one day than my whole flotilla normally produces in a year). I currently blame "something" about the second 570, as when I pulled it the system instantly became stable and has remained so.
The card:
XFX RX 5700 DD Ultra 8G
Model: RX-57XL8LB36
Price I paid: US$310
Initial impressions: This thing is BIG. It is the thickest card I've seen (termed a 2.5 slot card) and much longer and heavier than my XFX RX 570 cards.
The card has a tiny switch resembling the famous "mining mode BIOS" switch on the XFX RX 570 cards I've used. A little Internet searching suggested that also on this 5700 card XFX might have configured the forward (toward the I/O panel) switch position for a power-efficient mode of operation. I took a chance and moved that switch to that position on first install. Thus all results I post here may be non-representative of typical 5700 card behavior.
My very first impression was astonishment at how very little more power the card was burning when running Einstein than had my RX 570.
My second impression is that nevertheless there is a very considerable performance gain over the 570. Thus I am seeing good power efficiency gain at the card level, and a very good system level power efficiency gain--especially if I am later able to add back in my healthy 570, and perhaps, a little later, make this a dual 5700 system.
I ran a single WU at 1X. As I chose a _2 task, I got instant validation, which was comforting. The elapsed time running 1X was 6:12. I started running 2X partway through the second task, and am currently running steady 2X with elapsed times running about 10:47. That is a nominal daily credit rate of 925,000. Of course, in real life, the productivity will be less, as I must expect at least 1% loss to validation, and as this system is my daily driver, there will be also loss to task interference, down time, etc.
As I have complained about my VII on fan noise grounds, I should mention that this one is so far running the GPU fan at 33%, or 1153 rpm as reported by GPU-Z. It is not very noisy, although there is a noticeable surge at the end of a task. The character of the fan sound is quite pleasing (not a whine, beat tones, ... --think Noctua-like).
It is not only early days, but early minutes (and AMD has only had a driver release which would give Einstein GRP validations very recently).
I consider all of these results very preliminary, subject to adverse revision.
Current operating conditions (as reported by HWiNFO, and confirmed with GPU-Z)
AMD driver version: 20.1.2 released on January 20, 2020
GPU clock 1736 Max, 1707 Average
Memory clock 1750 unchanging
GPU core voltage 0.956 nearly always
GPU Temperature 76 C average
GPU Hot Spot Temperature 88 average
GPU fan 1810 peak, 1167 average
The approximate PC power draw at the wall socket is 225W--I'll take a good long-term average measurement later. For comparison, my (last night) good long-term average measure running a single RX 570 card at -35% power limitation in the same box was 188W!! At that condition the indicated daily productivity with the RX 570 was only 490,000.
If this holds together, I am very pleased. As the box has a Seasonic 850W supply and lots of fans, I'm also pretty optimistic about getting another card running in it (at least a 570, quite possibly another 5700).
I suspect the very low operating voltage is a big factor in the good power efficiency I am seeing. I don't know how much of that is because it is a 5700, and not a 5700 XT, how much because of the BIOS switch in "mining" position, or something else.
Archae86 wrote:I started
)
Congrats on your new card's success! I'm liking everything you've posted about it so far. I've been seriously considering getting that same card. Let us know how 3X works. And 4X!
Has anybody had success running gravity wave work on the 5700-series?
Ideas are not fixed, nor should they be; we live in model-dependent reality.