VCCSA above 1V causes instability

[WillyK]

R4E, i7-4960X, 32G RAM (DDR3-1600 XMP), BIOS 4901. Running 4.7G OC.

Auto VCCSA varies between 0.921-0.940V by default.
Too low, I thought. Bumping VCCSA a little should increase the IMC stability has been said.

However, If I try to bump VCCSA to values between 1.0-1.2V, the system gets unstable (Prime95 errors in 3-5 minutes).

How come? This is the opposite of what was expected and I don't understand that. Any ideas?

[WillyK]

Ok. It's interesting that I'm experiencing a particular bottleneck around 32K-36K FFTs in Prime95 while larger FFT are passing fine. E.g. I have another particular Prime95 test @ 1344K-1792K FFTs that has been a reliable detector of RAM flaws in most cases, and this one is running fine here... and it's pushing more data around than the 32-36K FFTs. Unfortunately, I don't have the knowledge required to understand why this might be happening on R4E with this CPU and this RAM and BIOS 4901.

[Raja]

The pattern of data used affects the current swings over the bus and that can impact timing margins - some patterns are tougher to pass based upon where the processor and memory signal stages are comfortable. As an example 11111111 may be easier than 1011011 as the latter is swinging between VOL and VOH while the latter slews to VOH and holds it for the duration of the burst.

[WillyK]

Good point. Thank you! I thought P95 varies the patterns quite a bit, but there may be something special with the 32K, 35K and 36K FFTs this test is running. It's a long shot but I'll check it out.

Notice that I'm not trying to break any records here. This is a very conventional RAM that most boards and CPUs out there would be happy to work with. Kingston is well known for manufacturing very reliable RAM and my kit has been up and running from day one. Since overclocked i7-4960X can allegedly cope with many RAM kits up to 2400-2600G, also on R4E, then working flawlessly @ DDR3-1600 on my R4E should have been walk in the park, right?

[Raja]

I think the lesson you should take from this it that when overclocked, things should not be assumed as being predictable or certain. CPUs can and do vary in what they are comfortable running. Memory kits are binned at stock processor core frequencies - if the bin is tight, the IMC weak with a particular IC, then it gives the potential for instability to manifest at a certain point.

There is nothing strange about the situation you are in believe me.

[WillyK]

So true. I'm a firm believer of the "cause-effect" relationship, always trying to find the logical reason behind, but this context is just too complex.

Anyway, here's the answer I got from the Mersenne Forum today:
"The only thing I can think of is that the 32K FFT starts using the L3 cache. Below that the majority of data will fit in the L2 cache.
That said, I don't see why the next larger FFTs wouldn't also have occasional trouble. But then again, Haswells are finicky beasts.
Sorry I can't be more helpful."

That makes a very interesting point but I'm not sure how to use this information.
Here's my CPU cache benchmarks:

[sandro_c]

Hi WillyK, hi Raja,
if my results can be of help on this tread I can communicate to you with my details overclocking:
Vcore v .............. 1,380 offset + 0.040
Vccsa v. ........... 1,090
Vtt v ................. 1,125
Ram v. .............. 1.50
multi ................. 44x in the summer , due to high room temps instead of 45X
BClock .............. 103 Mhz
Corefreq............ 4540 Mhz instead of 4648


I have not encountered instability with Vccsa about 1,150 v, only slightly more memory good results, and my bee it is not worth using that voltage


Greetings from
sandro c.

[catch22alex]

WillyK,

I'm experiencing the same issues as you with my 4930k. It seems that, for me, VCCSA below 1.0V produces the best results to keep my ram at 2133mhz

I went the same route and kept increasing with poor results, as the XMP profile on my ram sets the VCCSA to 1.25V

[WillyK]

catch22alex wrote:

I'm experiencing the same issues as you with my 4930k. It seems that, for me, VCCSA below 1.0V produces the best results to keep my ram at 2133mhz

I went the same route and kept increasing with poor results, as the XMP profile on my ram sets the VCCSA to 1.25V

Thanks for sharing! It means a lot to me to learn that I'm not the only case of the sort. I did a lot of testing and for my setup VCCSA=0.975V with VCCSA LLC=High and CC=120% works best, but ... still not fully stable (just it takes longer time to crash 1-2 Prime95 workers). I've never had setup so sensitive to the VTT/VCCSA values before (+/- 0.005 changes seem to make big difference).

Now, since each of the CPU and the RAM in itself are not unstable and indicate having even more headroom, I'm thinking that the only thing in between those two is the BIOS. The same RAM on the same MB did DDR3-1600 XMP without any trouble @ 4.6G OC but with SB CPU (3960X) and different BIOS! Now, this CPU sample (4960X) is actually better than my previous CPU and I have better water cooling as well. I know, R4E was not designed for IB-E so the BIOS is the key in my opinion.

I've experienced several times before, that after spending days on attempts to stabilize the system, just upgrading the BIOS with the brand new version just released by Asus would suddenly make the same system completely stable, without me changing anything! It has happened more than once.

Of course, I have no hard evidence proving my "BIOS theory" apart from the Auto values I'm observing that not always seem to be close to the optimal. Other people seem to be getting high OC with BIOS 4901 on R4E so maybe it's not the BIOS after all. Still, my guess is that next R4E BIOS version for IB-E may show much better stability with respect to the CPU-RAM dance.

[catch22alex]

WillyK wrote:
Thanks for sharing! It means a lot to me to learn that I'm not the only case of the sort. I did a lot of testing and for my setup VCCSA=0.975V with VCCSA LLC=High and CC=120% works best, but ... still not fully stable (just it takes longer time to crash 1-2 Prime95 workers). I've never had setup so sensitive to the VTT/VCCSA values before (+/- 0.005 changes seem to make big difference).

Now, since each of the CPU and the RAM in itself are not unstable and indicate having even more headroom, I'm thinking that the only thing in between those two is the BIOS. The same RAM on the same MB did DDR3-1600 XMP without any trouble @ 4.6G OC but with SB CPU (3960X) and different BIOS! Now, this CPU sample (4960X) is actually better than my previous CPU and I have better water cooling as well. I know, R4E was not designed for IB-E so the BIOS is the key in my opinion.

I've experienced several times before, that after spending days on attempts to stabilize the system, just upgrading the BIOS with the brand new version just released by Asus would suddenly make the same system completely stable, without me changing anything! It has happened more than once.

Of course, I have no hard evidence proving my "BIOS theory" apart from the Auto values I'm observing that not always seem to be close to the optimal. Other people seem to be getting high OC with BIOS 4901 on R4E so maybe it's not the BIOS after all. Still, my guess is that next R4E BIOS version for IB-E may show much better stability with respect to the CPU-RAM dance.

I should point out I have a R4BE, and it happened on multiple different BIOSes. From my perspective it has to do with the CPU itself. I'm happy where I'm at -- 4.4ghz @ 1.295Vcore and 2133mhz ram. 4.5 runs stable at 1600mhz RAM and I may do a bit more tweaking, but I know after 4.5 I need a ton of VCore to get to 4.6, it isnt worth it for me.

I agree, this is the hardest overclock I've had to do, but in the end it's because we have RAM that was tested at stock, and the IMC of each processor seems to behave differently when OCed.

[Raja]

You are on the right course of thinking Alex 🙂