Haswell-E Stability Testing

So, I thought Prime95 was the de facto stability test for Haswell-E - but after reading this article by ASUS ROG where they state:

Stress Testing
Users should avoid running Prime95 small FFTs on 5960X CPUs when overclocked. Over 4.4GHz, the Prime software pulls 400W of power through the CPU. It is possible this can cause internal degradation of processor components.

We use the ROG RealBench stress test, with the corresponding amount of installed memory selected in app. The stress test is equal or greater to alternatives, using real, open source apps with an oscillating load across the main PC subsystems.

Click to expand...

I am starting to question Prime95. I have been testing ROG RealBench and it seems to work very well at stressing all components of a system (not just CPU and/or memory alone). It also "errors" out in a lot more graceful manner. With Prime95, I'll set my stress test and walk away to find my PC completely powered down. I'm guessing it's the CPU shutting off due to it hitting the TJMax (my system configuration details are here).

What are your modes of testing stability with Haswell-E and have you tried ROG RealBench? Your thoughts on it?

I also did some further research (after ROG's claim in the above linked article that an 4.6GHz OC'd 5960X can pull 400W) regarding power usage by the 5960X when OC'd - it was extremely eye opening. Check these out:

Haswell-E Investigated: Overclocking and Power Consumption

Intel Core i7-5960X Underwater: Can a Custom Loop Raise the Ceiling?

Now I'm kind of worried about my Seasonic X-1250 1250W PSU being able to handle an OC'd 5960X plus Tri-SLI GTX 980s. I went to the power experts at the jonnyguru forums for that - thread here.

I think it is bullshit that Prime95 crashes a system like that. Just seems like poor design and overclocks that are too strong. The system crashes due to the AVX instructions putting too much stress on the PC and I believe the AVX instructions also cause a voltage increase in the CPUs.

The only time I'd consider this acceptable is if Prime or LinX crash the system due to excessive thermal load, not from a processor error. Most consumer-grade tasks don't draw nearly as much current as Prime - FFT multiplication is good at saturating the the core thanks to its predictable memory access patterns and data that fits in L3 cache (a 10M digit number is ~4MB). Unless you are doing FEA simulations or other similar scientific tasks it is unlikely you will generate the same amount of heat as Prime95 or LinX.

That being said, I'd consider BSODs and the like unacceptable during load testing; those are a sign that your CPU is incapable of generating correct results under load, never a good thing for stability.

Do you guys worry about a PSU with 25A on the +3.3V rail? Apparently, Raja from ASUS ROG recommends 30A (source) because an OC'd 5960X touches 25A on its own. I have some badass PSUs over here (Seasonic X-1250 and EVGA NEX1500 Classified) that have 25A on +3.3V. Gonna try the Corsair AX1500i.

---

dr/owned said:

I hit 30A on 1 rail and 20 A on the other. I presume you mean 12V because 3.3 is never loaded. I'd bet if I grab the wires they'll be hot at full load.

Click to expand...

That's all going to your 5960x? 600 watts?

I initially tried Prime95 when I first built my system but noticed that particular issue which made the CPU become a small reactor. Even though the CPU would work itself up to insane temps, the system would never freeze or reboot.
For stability testing now I use Handbrake and (re)encode a few dvd backups. This is pretty CPU intensive and will stress the system far more than any game or many other activities. I let this go on for an hour or 2.

---

dr/owned said:

I hit 30A on 1 rail and 20 A on the other. I presume you mean 12V because 3.3 is never loaded. I'd bet if I grab the wires they'll be hot at full load.

Click to expand...

Well, from your own PSU's specs (which, btw I ordered today ) it's max 30A on the +3.3V - but yeah I could be saying/interpreting things right because I am by no means a power supply expert.

I've been testing my 5960X all day for stability with overclocks (starting at 4.5GHz, 1.3v, then 1.325v, then 1.35v, then 4.4GHz, 1.3v, then 1.325v, and so on) - I'm now on 4GHz and still not stable with ROG RealBench. This test stresses ALL components (even my 3x980s) so my worry is my Seasonic X-1250 is just not providing enough juice. There's no way my 5960X is that much of a dog as to where I can only OC to 4GHz (perhaps lower, lol). It's cooled with a H110 and temps are quite good. Just wondering if it's that 25A that ASUS stated is a potential limit if you are OCing 5960X.

---

dr/owned said:

I think most PSUs will just click off if you exceed the current limit. The X-1250 is a fake single rail PSU. It actually has 4 12V rails with 30A/30A/45A/45A current limits, so there's a chance you're exceeding the 30A limit especially if your mobo only has a single 8 pin power connector. AX1200/1500i are true single rail.

Click to expand...

Any reference for this? I'm just referencing the load table. Same table exists for yours and outlines the 30A limit on 3.3V. I think what you're referring to is OCP. Which you can turn off with your unit (huge advantage). Pretty sure the X-1250 is not "fake" single rail.

---

dr/owned said:

Sorry, but why do you keep saying 3.3V? 12V is the important rail because that's what powers graphics cards and the CPU which are by far and away the biggest loads. I don't even know off the top of my head what uses 3.3V. Maybe various logic things on the mobo?

http://www.hardocp.com/article/2012/11/13/seasonic_x1250_1250w_power_supply_review/2#.VO_xQuH6ezy

See the "Alternate" header. I found out this fakeness was the case because back in the day I owned a X-1250 that would click off when running Furmark. Heavily overclocked 3770K + heavily overclocked set of 680 Lightnings (that I still have) and the X-1250's rails couldn't handle it. Same thing happened to the XFX 1000W PSU too because it's similarly a Seasonic design that's falsely advertised as single rail.

Click to expand...

Interesting! Thanks for the link.

I'm referencing the load table. The link that I sent earlier for your PSU (and soon mine) is here. I keep talking about the +3.3V because ASUS ROG calls that out as a possible issue for PSUs with an OC'd 5960X. If 25A - it's near the limit. If 30A - much better. See my above posts for reference for this.

---

dr/owned said:

From your own link (https://docs.google.com/file/d/0Bz2VRRbLPrZnYmJVVHM2UGxVS00/preview):

"PSU Requirements

For overclocking 5960X processors, we recommend PSUs that can supply 30 amps to EPS 12V. At 4.6GHz a 5960X can draw close to 25amps from the EPS12V connector at full load."

Click to expand...

Thank you - I told you I get confused with this power supply stuff...I was using this source and going off an article they referenced, but I think they were confused as well.

EDIT: they were - as it was resolved in the thread.

dr/owned said:

I was a bit off on my numbers. About 45A total going to the 5960X. And this is why I say Asus was being lazy / stupid / cheapskates in only giving the RVE 1x8 pin and 1 x 4 pin. There's Z97 boards with 2x8 pin for fuck's sake and they don't have 8 cores to deal with. "Extreme" would be 2x8 pin and a 4 pin molex or 6 pin PCI-e booster plug that adds even more power to the CPU.


free photo upload

Click to expand...

More pins wouldn't "add even more power to the CPU."

The CPU is powered by a switching regulator. When the motherboard manufacturers advertise "8 phases," for example, they mean that their switching regulator consists of an 8-phase switching regulator that steps the incoming 12V down to whatever voltage the CPU is using.

The keyword here is "regulator". The regulator will regulate the incoming voltage down to the CPU voltage over a wide range of input voltages. Note that this also means that the incoming voltage can vary quite a bit, because it's still going to be regulated down to the CPU voltage whether the input voltage is 11V or 12V.

Adding more connectors and wires between the PSU and the motherboard won't change that regulation on any significant scale. Adding more wires will only alleviate some of the voltage drop at the input of those regulators.

dr/owned said:

I hit 30A on 1 rail and 20 A on the other. I presume you mean 12V because 3.3 is never loaded. I'd bet if I grab the wires they'll be hot at full load.

Click to expand...

That power supply uses 16AWG wire for the big connectors, right? Given that the 30A load is going to be spread across a minimum of three 16AWG wires and you have some airflow through your case, I would guess that they're not actually hot at all.

BTW, I love that monitoring setup. Awesome power supply.

The Prime95 debate comes down to two camps:

1) Prime95 is a really, really good way to fully stress your system. So good, in fact, that you're unlikely to hit that same level of stress during real-world use, and you certainly won't hit it in any video game. As such, some people don't believe it's necessary to make your system fully stable for Prime95 and are willing to sacrifice this edge-case stability in the interest of a higher overclock that will probably work for the vast, vast majority of their time using the system.

2) Prime95 is the worst-case scenario for your system, so if you can make your system stable in this edge case then you know that it's going to be stable under the lesser workloads you experience elsewhere.


As for the idea that Prime95 will somehow degrade your CPU, I think that's a case of someone drawing an arbitrary line in the sand. AFAIK, there is no reason the CPU will suddenly stop throttling itself when the AVX instructions are used. So then the conjecture is that the increased current draw will somehow degrade the processor in ways that 'normal' overclocking somehow will not.

However, that idea doesn't really hold water, because most of the people who refuse to use Prime95 will just use the extra overhead to turn their clock speed and voltage up higher, which will, of course, just result in more current drawn and power dissipation in the processor.