RX 480 Results
30 Jun 2016 21:24:51 UTC
Topic 198669
(moderation:
My XFX RX 480 arrived today. After some initial struggles in which I learned there are as yet no functioning Windows 8 drivers (!), I've installed it onto my E5-2670 Windows 10 machine. It's running 3 concurrent BRP6 tasks; the system is also running CPU tasks from other projects and is hyperthreaded, so the individual cores are not terribly speedy.
Based on % complete per tick, it looks like it's going to come in just about exactly as fast at BRP6 tasks as a 7970 -- assuming this card is not afflicted by the validate error problem of the previous generation when running more than one simultaneous task.
I'll let it run 3 sets of 3; if all 9 tasks validate or are awaiting validation when those sets finish, I will keep it running 3x until I get a validate error, after which I will swap to 1x.
I'll get power numbers after the first set of 3. I forgot to plug in the Kill-a-Watt while I was grumbling about the Windows 8 issue.
RX 480 Results
)
That is fair. I'm eager to see your results. Does GPU-Z or something like it let you see what fraction of TDP it reports itself as consuming when running Einstein 3X?
Quick note, running out the
)
Quick note, running out the door: first set of 3 will be reporting abnormal times. I checked on the system to find all three still "running" but not updating. HWMonitor reported GPU usage as a flat 100%, whereas before it had been bouncing between 86 and 99. Suspending most CPU work got them progressing again.
I'm going to let these finish, power down, hook up the Kill-a-Watt, get some temp readings, drop to 2x and will try again with 0 CPU tasks.
First three BRP6 pending
)
First three BRP6 pending validation; again, times should be ignored because they were "running" but not really for an indeterminate number of minutes.
- System is dual E5-2670, two 115 TDP Xeons.
- Platinum power supply.
- Draw idle at wall with card, ~130w.
- Running two simultaneous BRP6, no additional CPU tasks:
- ~295w at wall. GPU-Z reports 100w drawn by card.
- ~85% (HWMonitor) / 89-98% (GPU-Z) GPU utilization. GPU-Z reports 50-60% memory controller usage and only 500MB of 8GB RAM used.
- Temperature is 81c-82c and 55% fan speed with ambient room about 24c. The card seems quiet.
- Note that the latest version of GPU-Z from 20 June is required to support the card.
- Note that the XFX has a very small factory core overclock (1288). This won't change results much.
One happy note; GPU-Z reports the ASIC score of the card is 86%.
RE: - Draw idle at wall
)
It seems to me that GPU-Z should be reporting something closer to 165 watts (i.e., 295w - 130w). But I have noticed that myself: GPU-Z shows a lower power usage (based on TDP) than I actually measure using a power meter. In my case, I usually use the power meter on my UPS (CyberPower CP1350PFCLCD), but when I use a Kill A Watt, I see about the same as the UPS. So I am wondering what GPU-Z is missing, or what else might account for it.
RE: RE: - Draw idle at
)
The GPU-Z number is probably accurate in that it references the power draw of the card itself (which has a maximum theoretical power of 150w, 75 from PCI bridge and 75 from 6-pin connector). The "missing" 65w would come from increased CPU and fan usage while the card is running and generating heat, as well as a 10% or so loss from power supply inefficiency.
RE: The GPU-Z number is
)
Nope,
GPU-Z relies on self-reporting from the card, and is described by the author as at best giving just the consumption of the GPU die itself and not the power consumption of the RAM and the power conversion circuitry on the card. I don't know whether he is well informed or not, and that is a second-hand report.
I, personally, think the reporting often inaccurate even for the card, and generally low. I've seen 30W reported for a 750Ti running Einstein, where other means suggest was burning rather closer to 50W.
People actually properly instrumenting the power consumption of the RX 480 card in reputable test sites have widely reported it to consume more than the 150W on a time-average (not just spikes) basis in some non-overclocked gaming and benchmark situations, and specifically to be exceeding the spec limits on both the motherboard and the 6-pin PCIe power connector. I have hoped that an RX 480 running Einstein might considerably underconsume one running modern games and benchmarks. However I think it implausible that merely servicing two RX 480 GPU tasks would drive up off-card power consumption by so much as 60W. An E5-2670 is a Sandy Bridge and not so very bad at power efficiency.
It is interesting that the card is reported as an Ellesmere device.
Two BRP6 sets of 2x now
)
Two BRP6 sets of 2x now complete. Third set in progress now has hyperthreading off, no other tasks running.
No results from wingmen yet.
Back to 3x now, and going to
)
Back to 3x now, and going to bed. To recap:
- 3x hyperthreaded, inaccurate(?) results from frozen(?) tasks.
- 2x, hyperthreaded
- 2x, hyperthreaded
- 2x, no hyperthreading. Slight improvement.
Now running 3x, no hyperthreading. GPU-Z is reporting peaks of 100% GPU usage. Power draw at wall up another 20, to ~315.
Please try the latest Beta
)
Please try the latest Beta version of HWiNFO: http://www.hwinfo.com/download.php
Not sure what exactly GPU-Z reports as power, but HWiNFO will report power consumed by the GPU Core (excluding other rails) and also should be able to report temperature and power from the GPU Core VRM. I don't think it's possible to measure the whole GPU power usage using software (at least for now).
Moreover it will allow you to monitoring GPU Memory Controller Errors - this is a unique feature no other tool offers and might be handy to determine if you're pushing the GPU memory above its limits.
There might be a few glitches in the current Beta regarding Polaris support (i.e. memory size reporting), this is already being fixed.
FYI - GPU Power reporting
NVIDIA offers a uniform method to report GPU Power via their API, but we don't know what's lying underneath it and what exactly does it include. However my comparisons with a meter have shown this pretty accurate.
AMD doesn't offer such API, so we have to use other methods - we can either directly report the power provided by a particular VRM (In/Out) in case the VRM/PWM is digital (+SMBus/I2C). Later AMD GPUs also have their own telemetry which they use to measure current/power from various rails. However in both cases we can only monitor a part of the entire GPU, mostly core + AUX or MEM rails. The rest is unknown as it doesn't feature any dedicated circuits for current measurement.
-----
RE: I, personally, think
)
I am inclined to agree with that. The other factors really can't add that much power. The DRAM memory for example refreshes itself at the same rate whether it is being used or not, and so its power will be relatively constant. The CPUs could add maybe 5 or 10 watts per core at most, but you are using only one core to support the GPU. And the 10% power supply loss would be only 15 watts for a 150 watt card. So I think that GPU-Z just does not measure everything on the card, and I won't use it so much for absolute power measurements, though it still may be useful for relative changes.