[Article] Not AVX Stable @ 4.4GHz, are we now? Shame!

I was able to run Intel XTU at 4.4GHz for 12Hrs+ and I thought it was stable. How wrong I was! It crashed in Handbrake after encoding videos for many hours. Once after 10hrs! After that I figured that there was something that I must have overlooked. When I asked around, I came to understand that XTU doesn’t use AVX in its stress test. Handbrake does, but still it doesn’t bombard it with AVX instructions like Prime95 does. BTW, Prime 95 27.9 (which is the latest version out with AVX support) crashes easily with the settings that XTU ran for hours and hours without failing.

Then I figured that it could be stable everywhere but AVX. One way to check it was to run Prime95 without AVX. That’s version 26.6. I ran it for hours and it didn’t fail. Now I was sure that it was the AVX instructions that caused the issue. 1.285V just isn’t enough when the components that makes AVX instructions work also become alive inside the CPU. I’m sure I would probably be able to escape by using Adaptive Vcore which automatically increases the Vcore by 0.1V when AVX instructions are heavily used. I tried 1.285V Adaptive, but it still failed in Handbrake. Probably the voltage fluctuation is also not a good thing.

Since then I’ve come to accept that you cannot have everything you want in life and 100MHz isn’t worth that increase in Vcore and temps (and increase in fan speed). So, now I’m going to settle with my 4.3GHz OC and tweak it a little bit to get every bit of power I can get with it. I’ve successfully managed to up the Uncore to 39x at just 1.08V and Memory to 1600MHz @ XMP. I’m also in the process of tweaking the Digi+ power delivery settings (image below) to lower the power consumption and the temps as much as possible without crippling stability. The following screenshot was taken while running Handbrake. Look at the fan RPM – bottom right hand side. It’s completely silent. That’s just great. I won’t be able to achieve that with close to 1.3V. The temps it shows is not very accurate though. Those are not the core temps. Core temps hover between 60C and 70C.

I haven’t done a lot of testing, but after dropping the LLC to level 6, enabling Active Frequency Mode and setting CPU Power Phase Control to optimized, at least it is stable in Intel Burn Test, Handbrake and Crysis 3. If it passes 6hrs of Prime95, I say I’m stable enough.

[Article] Haswell stress testing using Intel Extreme Tuning Utility(XTU)

I'm still trying to find the sweet spot for my Haswell CPU overclock. Prime95 has been failing on my in a weird way. Sometimes it would go for hours without failing. Sometimes even a slight voltage bump would cause it to crash instantly. If I go back to the settings that worked fine earlier, it would crash again.

Last night I decided that it was high time that I moved away from Prime95 and found something new that worked well with Haswell. I have tried AIDA64, but people have been complaining that even after passing it for 12Hrs, it would crash in games. I don't know if that is a good stress test either. Plus, you have to pay for it.

Some guy one Overclock forums recommended that he is using Intel's own stress testing utility called Intel Extreme Tuning Utility (a.k.a Intel XTU). People say that it fully supports the new AVX2 instructions, doesn't make the Vcore go out of control when the CPU is bombarded with AVX instructions and it also shows all the voltages and temps on the same screen (so that you don't have to run a separate monitoring utility (or utilities) as is the case with Prime95. AIDA64 also shows the voltages and temps while stress testing. However, XTU also allows you to overclock the CPU from itself. But I don't really like overclocking using software, especially when AISuite III is installed there and it can clash with other software.

So, I started a stress test in XTU last night and I ran it for 13hrs without a crash or error. However, I'm not confident that not failing XTU alone is a good indication of stability. So I am running a few hours' long H.264 transcode queue in HandBrake. If I can get through it without an error, then I guess I would be stable in everything that I do. I only stress tested the CPU portion. I skipped memory for now. If H.264 fails, I would test Memory as well. Otherwise I don't see why I need to do that because the clocks, timings and voltage is pretty standard for today.

The current overclock settings stands as follows.

• Core multiplier: 44x (sync all cores enabled)
• Vcore: 1.285V manually set in BIOS (should change to Adaptive later)
• Un-core multiplier: x39 (AUTO)
• Un-core voltage: 1.15V (AUTO)
• RAM speed: 1600MHz (XMP)
• RAM timings: 9/9/9/24/2T (XMP)
• RAM voltage: 1.50V (XMP)
• Input Voltage: 1.75V (AUTO)
• Extreme Tweaking: Disabled
• EPU Power Saving Mode: Disabled
• Everything else: AUTO

※This is on hte Asus Maximus VI Hero board with BIOS ver.711.

Here's proof.

Note: 
Time remaining says 10hrs 58min because I set it to 24hr stress test. Deduct that from 24hrs, and you get 13hrs 2min. 

[Article] Haswell overclocking woes - the deal with Vcore, IVR and AVX instructions

In the previous article I discussed what the main causes of high temps on Haswell CPUs. Now let's look at another issue.

When you overclock your CPU, there are two main method how you would set the Vcore. Static mode and Offset mode.

Static mode

Static mode means that even when the CPU is idle or doing light work, the Vcore would remain as if you are doing heavy work. The Vcore would not change dynamically according to the work load. Now, there is no real need for that. Lower Vcore is always better for longevity of the chip.

Offset mode

Thus there is Offset mode. Intel CPUs automatically can set a Vcore based on the multiplier by referring to voltage tables embedded in the chip. These voltages are called VIDs and higher the multiplier, higher the VID. But Intel can only guarantee the VID is stable only within the specifications. After that, if can be either stable or not. At some point you'll find out that the stock VID is not stable enough. Thus you need can specify an offset to the dynamic VID. It doesn't have to be positive. If you know your CPU can work at a lower Vcore than the VID at max multiplier, you can specify a negative offset. But there is one issue. The offset applies to entire multiplier range. That means, even at idle, the offset value would be applicable. Especially if you are using a negative offset, it can crash the system. Or it won't be optimal.

So with Haswell, Intel introduced a 3rd method. Adaptive mode.

Adaptive mode (Haswell only)

Adaptive mode works like Offset mode, but only applies to turbo multipliers. That means, the dynamic Vcore would be identical to the VID below turbo frequencies. Thus, this method does not cause instability at idle states due to negative offsets an does not over-volt within the specified multiplier. Pretty good deal.

Except, it is not. Actually, it's not Adaptive mode's fault.

I'm sure that you already know that there is an IVR onboard the CPU package. The IVR is not the entire CPU voltage regulator. The old VRMs are still there on the board and are needed. The IVR is there to control the Vcore more aggressively. Basically, the motherboard inputs the CPU a quite large voltage called the Input Voltage. The IVR give the cores a portion of the this Input Voltage. It can rapidly fluctuate voltage depending on the load and multiplier.

OK, now that looks like a nice thing. But AVX instructions make the Vcore go crazy. I don't know why, but for some reason, whenever the AVX instructions are used, the IVR increases the Vcore by about 0.1V. This happens in Offset mode as well. And there is nothing you can do about it other than use static Vcore. With static Vcore, you don't have this problem. Even when AVX instructions are being used, the Vcore doesn't change.

Stress testing with AVX loads is the problem

This doesn't look as bad as it sounds in reality because you might not use AVX instructions at all. Plus, the increase in Vcore doesn't necessarily increase the temps in normal work loads as if you are running a stress test. If you run Prime95 or LinX or AIDA64 which stresses the CPU with AVX loads, it would definitely make the temps skyrocket. For this reason, they say not to stress test using Offset or Adaptive methods. Old Prime95 (pre 27.7) is OK because it doesn't support AVX. Now, one option would be to run at Offset mode or Adaptive mode and then switch to static mode once you start using AVX supported applications, if you ever. The upcoming H.265 encoding is supposed to make use of AVX instructions, and that would be a problem for Adaptive mode or Offset mode. I'm hoping to move to H.265 once it becomes stable. You get the same quality for half the file size. That's something you don't wanna miss out on.

There shouldn't be a big difference in power consumption between static Vcore mode and either of the dynamic Vcore modes as long as you are using all the power saving features. In fact, some say (I haven't been able to confirm it) that on the Asus Z87 boards, at full idle, the Vcore drops to idle Vcore levels as if it is using Offset mode. This seems to be how Gigabyte boards operate. I'm gonna check this out when I get back home. I think the board has the C6/C7 power states disabled by default. My PSU should be able to handle those low currents.

My 4.5GHz overclock on 2600K has become unstable!

It's been a while since I tested for stability of my CPU's 4.5GHz overclock. I ran the blend test of Prime95 of more than 6hrs back then and I felt that was sufficient. However, PC has been giving random lockups in the past couple of months and I didn't want to accept that the reason for this instability was because my overclock wasn't stable enough. For some reason, it was so hard to find the settings that made my PC stable, and I didn’t want to go through all that again. It could be because I wanted to both overclock and save power at the same time. That's not something that's easy to achieve, obviously. Or it could be because the motherboard is not the best when it comes to overclocking. Or it can even be that this CPU is a bad clocker. Or I could have been doing something wrong in the first place.

Anyways, the reason for the recent instability could be due to many reasons. The first one could be because the last time I checked for stability, I didn't use a version of Prime95 that supported AVX instructions. Then, I did a couple of UEFI updates and maybe I needed to change “some values” to make the CPU stable under the firmware that I'm using currently. Or it could well be a problem with the power supply or motherboard or memory.

 

Prime95 with AVX support for Stress Testing your Sandybridge or IvyBridge CPU

I discovered this recently. I’ve been using Prime 95 without AVX support for all this time. Prime 95 with AVX support is the bomb. Just compare the temperature difference between the following screenshots.

Note: Small FFT stress test was run for 5 minutes.

Without AVX (v26.6)