Anand had some Compute benchmarks in it's review. It did very poor in FAH which is all FP32. Blender and Compubench it was better than NV cards. Not sure how E@H would fit in there but I guess in the middle of those results maybe?
Yeah, I've seen those. I would expect it to perform somewhat better than Fiji at Einstein. But don't know whether they have fixed the issue of not being able to run more than 1 WU concurrently. If not there may be no advantage over Fiji at all (almost the some bandwidth). But let's wait for actual reports :)
Edit: 1st report and pretty fast (370 - 373 s for a single V1.18 WU), although not exceptional (my GTX1070 is doing 2 WUs in 1220 s, so effectively 610 s per WU - but at just ~100 W).
RX Vega 64 just arrived. First quick test run at default settings gives 480-510 s.
Something must be wrong as the Fury X does these in 460 s, will need to check and adjust..
EDIT: OK, so profile was set to Balanced. Now trying Turbo.
EDIT2: Turbo doesn't seem to improve it, similar results.
EDIT3: Custom profile, +15% power limit - same result.
EDIT4: Some units made it at 430-460 s, seems even in this small set of units, the work amount is variable, so not good for comparison.
During all these tests the BIOS switch was always in the primary (higher power limit) position and the max GPU clock was ~1280 MHz, usually it's ~1100-1200 MHz. Seems the GPU doesn't boost. Power usage is very difficult to estimate due to the nature of these tasks, where GPU load (and power) fluctuates extremely.
Result: Performance running FGRPB1 is comparable to Fury X.
For more comparison - units of this batch take ~700 s (1 WU) on RX 480.
EDIT5:
Now let's see how running multiple units works.
So 2xFGRPB1 gives ~620 s. Pretty good. Now the average GPU clock is higher (~1300 MHz) and it sometimes reaches the boost top (1540 MHz), so in the first test the clock wasn't boosting due to the highly fluctuating load, which probably wasn't sufficient to trigger higher clocks.
The "GPU Chip Power" value hits the TDP (220 W) pretty often now, so clocks are most probably reduced due to this. Raising power limit shows immediate effect including raised temperatures and time improve by some tens of seconds depending on the actual level.
Not sure yet if these units will validate, as running concurrent WUs was a problem with recent AMD GPUs.
Temperatures:
Running 1 WU the highest temperature reaches is 70-72 C, 2 WUs 75-80 C.
IMPORTANT NOTE: If anyone wants to make a comparison with other GPUs, note that the FGRPB1 tasks have a very variable workload/run time. So one needs to compare with results of the same batch.
IMPORTANT NOTE: If anyone wants to make a comparison with other GPUs, note that the FGRPB1 tasks have a very variable workload/run time. So one needs to compare with results of the same batch.
My GTX 1070 at stock speed (don't remember what it was, but it is not factory overclocked very much) under Ubuntu 16.04 takes just under 700 seconds. But the average power usage was well below maximum, though it varied a lot of course. I would guess it was about 110 watts, as compared to the TDP value of 150 watts.
IMPORTANT NOTE: If anyone wants to make a comparison with other GPUs, note that the FGRPB1 tasks have a very variable workload/run time. So one needs to compare with results of the same batch.
My GTX 1070 at stock speed (don't remember what it was, but it is not factory overclocked very much) under Ubuntu 16.04 takes just under 700 seconds. But the average power usage was well below maximum, though it varied a lot of course. I would guess it was about 110 watts, as compared to the TDP value of 150 watts.
Is this x1 ?
I see your last results returned on Sep-3, you might need to run some of the actual ones in hope to catch the same batch as I have run.
OK, some are running now. You can check them at your leisure. Yes, that is x1, which I prefer.
By the way, my preferences (both computing and Project) were seriously mixed up from the last time I changed them, only a few days ago. I think the server still has some problems.
Is this x1 ?
I see your last results returned on Sep-3, you might need to run some of the actual ones in hope to catch the same batch as I have run.
OK, some are running now. You can check them at your leisure. Yes, that is x1, which I prefer.
By the way, my preferences (both computing and Project) were seriously mixed up from the last time I changed them, only a few days ago. I think the server still has some problems.
The only way I have to measure power on that GTX 1070 is to look at the digital output meter on the UPS, an APC Back-UPS 1000. It jumps around a lot, but with the card operational the total power to the PC looks to be around 218 watts, and with the card disabled from running the Einstein GPU work (leaving only the Einstein CPU work) it is about 103 watts, or 115 watts for the FGRPopencl1K-nvidia (version 1.20) work on that card.
However, that is at the beginning of the run, and it drops to a lower value later. Also, that value includes the power supply loss. It is a gold-rated supply, about 90% efficient, so the actual differential power to the card is about 104 watts for that work. On the other hand, it does not include the static power to the card, which is probably around 10 watts, so I would say that 114 watts is what you would use to compare the Einstein FPRPG1 work on that card to the AMD Vega, or whatever else you would compare it to. If I took an average over the entire run, I think it would be even less, since maybe the CPU does more of the work later in the run as I remember.
Haven't spotted any here yet,
)
Haven't spotted any here yet, but my Vega is already on the way from AMD. Will report ASAP when I get it.
-----
Anand had some Compute
)
Anand had some Compute benchmarks in it's review. It did very poor in FAH which is all FP32. Blender and Compubench it was better than NV cards. Not sure how E@H would fit in there but I guess in the middle of those results maybe?
http://www.anandtech.com/show/11717/the-amd-radeon-rx-vega-64-and-56-review/17
Yeah, I've seen those. I
)
Yeah, I've seen those. I would expect it to perform somewhat better than Fiji at Einstein. But don't know whether they have fixed the issue of not being able to run more than 1 WU concurrently. If not there may be no advantage over Fiji at all (almost the some bandwidth). But let's wait for actual reports :)
Edit: 1st report and pretty fast (370 - 373 s for a single V1.18 WU), although not exceptional (my GTX1070 is doing 2 WUs in 1220 s, so effectively 610 s per WU - but at just ~100 W).
MrS
Scanning for our furry friends since Jan 2002
RX Vega 64 just arrived.
)
RX Vega 64 just arrived. First quick test run at default settings gives 480-510 s.
Something must be wrong as the Fury X does these in 460 s, will need to check and adjust..
EDIT: OK, so profile was set to Balanced. Now trying Turbo.
EDIT2: Turbo doesn't seem to improve it, similar results.
EDIT3: Custom profile, +15% power limit - same result.
EDIT4: Some units made it at 430-460 s, seems even in this small set of units, the work amount is variable, so not good for comparison.
During all these tests the BIOS switch was always in the primary (higher power limit) position and the max GPU clock was ~1280 MHz, usually it's ~1100-1200 MHz. Seems the GPU doesn't boost. Power usage is very difficult to estimate due to the nature of these tasks, where GPU load (and power) fluctuates extremely.
Result: Performance running FGRPB1 is comparable to Fury X.
For more comparison - units of this batch take ~700 s (1 WU) on RX 480.
EDIT5:
Now let's see how running multiple units works.
So 2xFGRPB1 gives ~620 s. Pretty good. Now the average GPU clock is higher (~1300 MHz) and it sometimes reaches the boost top (1540 MHz), so in the first test the clock wasn't boosting due to the highly fluctuating load, which probably wasn't sufficient to trigger higher clocks.
The "GPU Chip Power" value hits the TDP (220 W) pretty often now, so clocks are most probably reduced due to this. Raising power limit shows immediate effect including raised temperatures and time improve by some tens of seconds depending on the actual level.
Not sure yet if these units will validate, as running concurrent WUs was a problem with recent AMD GPUs.
Temperatures:
Running 1 WU the highest temperature reaches is 70-72 C, 2 WUs 75-80 C.
IMPORTANT NOTE: If anyone wants to make a comparison with other GPUs, note that the FGRPB1 tasks have a very variable workload/run time. So one needs to compare with results of the same batch.
-----
Mumak wrote:IMPORTANT NOTE:
)
My GTX 1070 at stock speed (don't remember what it was, but it is not factory overclocked very much) under Ubuntu 16.04 takes just under 700 seconds. But the average power usage was well below maximum, though it varied a lot of course. I would guess it was about 110 watts, as compared to the TDP value of 150 watts.
I have no idea of the batch. https://einsteinathome.org/host/12360156/tasks/0/40
Jim1348 wrote:Mumak
)
Is this x1 ?
I see your last results returned on Sep-3, you might need to run some of the actual ones in hope to catch the same batch as I have run.
-----
R290 under Ubuntu 16.04 runs
)
R290 under Ubuntu 16.04 runs 3x tasks in ~1320s, so ~440s/task
https://einsteinathome.org/host/12494837/tasks/4/0
RX480 performance is similar:
https://einsteinathome.org/content/best-cost-efficient-gpu-einstein?page=1#comment-155602
So OS matters.
Quote:Mumak wrote: Is this
)
OK, some are running now. You can check them at your leisure. Yes, that is x1, which I prefer.
By the way, my preferences (both computing and Project) were seriously mixed up from the last time I changed them, only a few days ago. I think the server still has some problems.
Jim1348 wrote:Mumak
)
OK, so these take 760 s now.
-----
The only way I have to
)
The only way I have to measure power on that GTX 1070 is to look at the digital output meter on the UPS, an APC Back-UPS 1000. It jumps around a lot, but with the card operational the total power to the PC looks to be around 218 watts, and with the card disabled from running the Einstein GPU work (leaving only the Einstein CPU work) it is about 103 watts, or 115 watts for the FGRPopencl1K-nvidia (version 1.20) work on that card.
However, that is at the beginning of the run, and it drops to a lower value later. Also, that value includes the power supply loss. It is a gold-rated supply, about 90% efficient, so the actual differential power to the card is about 104 watts for that work. On the other hand, it does not include the static power to the card, which is probably around 10 watts, so I would say that 114 watts is what you would use to compare the Einstein FPRPG1 work on that card to the AMD Vega, or whatever else you would compare it to. If I took an average over the entire run, I think it would be even less, since maybe the CPU does more of the work later in the run as I remember.