Why different versions of Prime stress differently

[Chesebert]

I was doing final stress testing on my 4790k and being an old fart I am still using Prime 25.x. My temps are in the low 70s and all is great. I then downloaded Prime 27.7 and 28.5 and both of which made my system scream. Prime on 27.7 takes me into the high 80s and 28.5 goes straight to high 90s and reaching 100!

All prime versions are run with small FFT test. What has changed?

I can run x264 on days end and the CPU stay in the 60s.

What's going on?

 

[Racan]

Prime 28.5 uses Haswells AVX2 and FMA3 instructions.

 

[Yuriman]

Most disable the AVX2 extensions in Prime when stress testing on Haswell, because they add voltage and are a non-typical use scenario.

 

[Nothingness]

Most disable the AVX2 extensions in Prime when stress testing on Haswell, because they add voltage and are a non-typical use scenario.

So they stress test by disabling a feature some other program might use? That isn't stress test any more

 

[Dufus]

AVX2 is a wider instruction, 256 bits vs SSE at 128bits and MMX at 64bits so when optimized for AVX a lot more switching can go on.

For instance running Linpack bench with AVX disabled on my laptop results in 53GFLOPS and uses 55W while at the same clock and AVX enabled results in 187GFLOPS and 84W. Obviously it also generates more heat / higher temps.

FWIW difference in voltage using AVX on my own processor running adaptive interpolation mode is ~10mV (0.010V), however there are reports of voltages increasing by 100mV (0.1V).

It'll be interesting when/if Cannonlake makes use of AVX512, 512bit instructions.

 

[ShintaiDK]

The main reason is that with 256bit AVX/AVX2, Haswell changes to 256bit cache width and speed. It adds +100mv while nearly doubling the throughput.

 

[Dufus]

Here, I've just run Linx with Linpack 11.2.1 to show Vcore and these are the results. Turbo has been set to a flat 34x so the 35x and 36x unlocked bins do not skew the results.

The voltage while running AVX2 is pretty close to when idle and cool and in fact drops when coming off maximum load. This seems to be temperature related. I don't know why others are experiencing a 100mV increase but it's not happening here.

 

[Nothingness]

Here, I've just run Linx with Linpack 11.2.1 to show Vcore and these are the results. Turbo has been set to a flat 34x so the 35x and 36x unlocked bins do not skew the results.

The voltage while running AVX2 is pretty close to when idle and cool and in fact drops when coming off maximum load. This seems to be temperature related. I don't know why others are experiencing a 100mV increase but it's not happening here.

Perhaps is it because you have a mobile chip and these are binned more tightly than desktop variants? My 4770K certainly raises the voltage by 0.1V when running linx.

 

[soccerballtux]

So they stress test by disabling a feature some other program might use? That isn't stress test any more

+1

 

[BSim500]

So they stress test by disabling a feature some other program might use? That isn't stress test any more

The "stress test" of AVX power viruses that abnormally trigger a +0.1v auto-overvolt on Haswell's rarely reflect the reality of normal usage. Eg, the most typical heavy constant 100% on all core load activity the average person will do is X264 encoding, and temps are regularly 20-30c lower on Haswell's vs Prime (even though X264 uses AVX instructions) purely from the voltage change that Prime, Linpack, IBT, etc, trigger off. You could accomplish the same thing on any chip - go into your BIOS and add +0.1v vCore on top of what you normally run it at. Hey presto - you're now "stressing" it the same way without the resulting artificially higher temps actually being of any use regarding how stable the CPU will be on normal apps / games.

 

[witeken]

That's also why the Haswell Xeons have different TDPs/clock speeds for AVX2 workloads.

 

[BonzaiDuck]

I was doing final stress testing on my 4790k and being an old fart I am still using Prime 25.x. My temps are in the low 70s and all is great. I then downloaded Prime 27.7 and 28.5 and both of which made my system scream. Prime on 27.7 takes me into the high 80s and 28.5 goes straight to high 90s and reaching 100!

All prime versions are run with small FFT test. What has changed?

I can run x264 on days end and the CPU stay in the 60s.

What's going on?

Tell me about your CPU cooling solution.

 

[Nothingness]

The "stress test" of AVX power viruses that abnormally trigger a +0.1v auto-overvolt on Haswell's rarely reflect the reality of normal usage.

prime95 isn't a power virus. Some people run it for months. Of course it isn't representative of what most people do with their PC, but then do you think it would make sense to run a browser to "stress test" your system?

Eg, the most typical heavy constant 100% on all core load activity the average person will do is X264 encoding, and temps are regularly 20-30c lower on Haswell's vs Prime (even though X264 uses AVX instructions) purely from the voltage change that Prime, Linpack, IBT, etc, trigger off. You could accomplish the same thing on any chip - go into your BIOS and add +0.1v vCore on top of what you normally run it at. Hey presto - you're now "stressing" it the same way without the resulting artificially higher temps actually being of any use regarding how stable the CPU will be on normal apps / games.

That's completely different: the +0.1V when AVX is used is just how Haswell works. And again we are not talking about "normal" usage, we are talking about stress testing.

 

[Nothingness]

That's also why the Haswell Xeons have different TDPs/clock speeds for AVX2 workloads.

Yes, reduced turbo speed when AVX2 is used. There is some nice explanation here: http://www.microway.com/knowledge-c...s-intel-xeon-e5-2600v3-haswell-ep-processors/

 

[Chesebert]

Tell me about your CPU cooling solution.

I have H100i installed. I figured out my problem was using adaptive voltage. I have set my VID to 1.26v, which gets raised to 1.28v during OCCT and Prime.

At 1.28v, I am getting low 80s during OCCT and Prime 27.7. Running x264 benchmark the chip stays in the high 60 and low 70s. Prime 28.5 is still very crazy and takes the temp into the 90s, but I don't think that test is all that relevant.

4790k @ 4.8ghz with 4ghz uncore.

 

[Dufus]

Yes, reduced turbo speed when AVX2 is used. There is some nice explanation here: http://www.microway.com/knowledge-c...s-intel-xeon-e5-2600v3-haswell-ep-processors/

Some nice details however...

While a CPU core is executing intensive vector tasks (AVX instructions), the clock speed may be reduced to keep the processor within its power limits (TDP). In effect, this may result in the processor running at a lower frequency than the “base” clock speed advertised for each model. For that reason, each “Haswell” processor model is assigned two “base” frequencies:

1. AVX mode: due to the higher power requirements of AVX instructions, clock speeds may be somewhat lower while executing AVX instructions *
2. Non-AVX mode: while not executing AVX instructions, the processor will operate at what would traditionally be considered the “stock” frequency

This doesn't seem to be any different to the way power limiting works with the last few gen's except it appears the "clamp" bit is being used to drop the multi below the HFM if required.