So it's widely known that the Haswell VRM will automatically boost vcore when running AVX instructions. Usually, this is regarded as a Bad Thing (TM) but I did some testing tonight that doesn't quite fit with this thinking.
Until now, my preferred method of overclocking has been a fixed vcore and speedstep turned off. At 4.4 GHz, I've found that:
1.24 is required for AVX1, e.g. a 12-hour Prime95 run or Aida64 FPU test
1.28+V is required for Linpack 11 with AVX2 (it works at 1.24V, but BSODs every couple of runs).
With non-AVX loads, I could probably use a much lower voltage, but since the VCore is fixed, I have to set it for a "worst case" scenario.
1.28+V results in 90+C in Linpack 11 and I also have to make a choice – do I stick with 1.24V which works fine in everything else, or do I put in a needlessly high voltage just so I can run Linpack 11?
Today I tried adaptive vcore for the first time and dialed in +0.030V.
At 4.4 GHz, this results in
At full non-AVX load (e.g. Aida64 CPU test): 1.197V
In AVX/AVX2 tests: 1.273V
However, running Linpack 11, the voltage actually only hit 1.273V during very brief periods of time. Most of the time, it was actually at 1.197V. This resulted in a Linpack AVX2 peak temperature of 76C.
I think it's doing the same thing in Prime95. My Prime95 temps are now actually 75C with Small FFT's, vs 80C at 1.24V fixed before, even though Aida64 indicates a constant 1.269V.
Thanks to the very low non-AVX voltage, the Aida64 CPU-test temps are now ridiculously low and topped out at 52C now at 1.197V.
So far I've only done 10 minute runs of each stress test, so I can't confirm 24/7 stability. But so far, it looks *very* promising compared to the temps I was seeing with fixed Vcore. It almost looks like you can use the AVX voltage boost to your advantage by using a lower non-AVX voltage. It looks like Intel very obviously designed AVX2 with dynamic VCore in mind, only applying the +0.1V boost for a fraction of a second when it is required.
Until now, my preferred method of overclocking has been a fixed vcore and speedstep turned off. At 4.4 GHz, I've found that:
1.24 is required for AVX1, e.g. a 12-hour Prime95 run or Aida64 FPU test
1.28+V is required for Linpack 11 with AVX2 (it works at 1.24V, but BSODs every couple of runs).
With non-AVX loads, I could probably use a much lower voltage, but since the VCore is fixed, I have to set it for a "worst case" scenario.
1.28+V results in 90+C in Linpack 11 and I also have to make a choice – do I stick with 1.24V which works fine in everything else, or do I put in a needlessly high voltage just so I can run Linpack 11?
Today I tried adaptive vcore for the first time and dialed in +0.030V.
At 4.4 GHz, this results in
At full non-AVX load (e.g. Aida64 CPU test): 1.197V
In AVX/AVX2 tests: 1.273V
However, running Linpack 11, the voltage actually only hit 1.273V during very brief periods of time. Most of the time, it was actually at 1.197V. This resulted in a Linpack AVX2 peak temperature of 76C.
I think it's doing the same thing in Prime95. My Prime95 temps are now actually 75C with Small FFT's, vs 80C at 1.24V fixed before, even though Aida64 indicates a constant 1.269V.
Thanks to the very low non-AVX voltage, the Aida64 CPU-test temps are now ridiculously low and topped out at 52C now at 1.197V.
So far I've only done 10 minute runs of each stress test, so I can't confirm 24/7 stability. But so far, it looks *very* promising compared to the temps I was seeing with fixed Vcore. It almost looks like you can use the AVX voltage boost to your advantage by using a lower non-AVX voltage. It looks like Intel very obviously designed AVX2 with dynamic VCore in mind, only applying the +0.1V boost for a fraction of a second when it is required.
Last edited: