OK, I did a quick test with your settings, but can't reproduce the results.
Undervolting the GPU clock to 1.025 V and +50% power limit, the GPU voltage reported still fluctuates between 1.05 - 1.1 V, GPU power reported is 180 - 220 W.
With just reducing the power limit to -20% the GPU voltage fluctuates between 0.881 - 1.175 V and power 129 - 178 W.
This is all running 2x WU. When running just 1 WU the results for undervolting look much better.
Many thanks indeed for posting all your testing results and for taking the time to honour my requests ;-)
I will be re exploring undervolting my Vega cards as soon as my time allows, likely not before the weekend.
One thing that is intriguing me is your runtimes v's mine, yours being noticeably longer yet with your reduced power consumption and temperatures your card is able to sustain higher core clock frequencies and has the healthy 55Mhz memory overclock... On paper your card should be faster (or at least equally as quick) whilst using less juice. This leads me to one more question.
Do you have numbers or even a general feeling for how undervolting effects runtime compared to running stock values?
Memory is hardly used afaik for Einstein WUs. Also - im running Ryzen 7 1700 (at stock), which peaks at around 3.2Ghz - you've got Vega64 on 4790K which for single thread tasks is way faster (one part of Einstein WU is almost purely ran on CPU, i think its validation of some sort).
I'm going to try HWinfo64 for logging (h/t @Mumak) and see if it catches something different.
@Gavin - regarding actual run times and if undervolting caused any changes - since i did run 2WU with both undervolt and at stock voltages in last test - there were no differences in run time between them.
(No subject)
)
Very nice, thanks! I will
)
Very nice, thanks! I will give this a try.
-----
OK, I did a quick test with
)
OK, I did a quick test with your settings, but can't reproduce the results.
Undervolting the GPU clock to 1.025 V and +50% power limit, the GPU voltage reported still fluctuates between 1.05 - 1.1 V, GPU power reported is 180 - 220 W.
With just reducing the power limit to -20% the GPU voltage fluctuates between 0.881 - 1.175 V and power 129 - 178 W.
This is all running 2x WU. When running just 1 WU the results for undervolting look much better.
-----
Which drivers?
)
Which drivers?
45 minutes of graphed timeframe, stats.
Latest - Crimson ReLive
)
Latest - Crimson ReLive Edition Beta for Windows 10 Fall Creators Update.
-----
Wattman Didnt work for me as
)
Wattman didnt work for me as expected with these, i reverted to 17.9.3 WHQL due to that.
Many thanks indeed for
)
Many thanks indeed for posting all your testing results and for taking the time to honour my requests ;-)
I will be re exploring undervolting my Vega cards as soon as my time allows, likely not before the weekend.
One thing that is intriguing me is your runtimes v's mine, yours being noticeably longer yet with your reduced power consumption and temperatures your card is able to sustain higher core clock frequencies and has the healthy 55Mhz memory overclock... On paper your card should be faster (or at least equally as quick) whilst using less juice. This leads me to one more question.
Do you have numbers or even a general feeling for how undervolting effects runtime compared to running stock values?
Gav.
Memory is hardly used afaik
)
Memory is hardly used afaik for Einstein WUs. Also - im running Ryzen 7 1700 (at stock), which peaks at around 3.2Ghz - you've got Vega64 on 4790K which for single thread tasks is way faster (one part of Einstein WU is almost purely ran on CPU, i think its validation of some sort).
I'm going to try HWinfo64 for logging (h/t @Mumak) and see if it catches something different.
ikeke1 wrote:I'm going to try
)
Use the latest HWiNFO Beta - it adds Hot Spot temperature and GPU SoC clock monitoring.
-----
@Gavin - regarding actual run
)
@Gavin - regarding actual run times and if undervolting caused any changes - since i did run 2WU with both undervolt and at stock voltages in last test - there were no differences in run time between them.