5960x 4.8 @ 1.3v / 5.0 @ 1.35v (not kidding)

[cekim]

Never hurts to go over your system with a fine-tooth comb... I found my PWM EK pump voltages were way off as I had tuned the wrong thing in the BIOS (I had hooked up the pump control to 4 and the fans for the radiator to 1 and then tuned 1 as if it were the pump).

That wasn't the only issue I found, nor can it fully explain this. Short run-down of issues and possible impact:
- water pump was nearly off/off at any temp short of 75C (and still remarkably effective - but temps are much narrower core-to-core now)

- all fans were running harder than they needed to be (there are 12 fans, 3 140mm, 3 200mm), meaning they were drawing current and inserting noise I am sure through the MB

- random PCIe conflicts with M.2 and 10GbE card (PCIe lanes all busy - SLI x16x2 + 10GbE + m.2). Solved by using hyper card instead of m.2

So, with that all out of the way:
1. ~10 minutes of aida64 at 4.8 @ 1.3v (it eventually found instability, but look at those temps?)
2. 5GHz club :cool:

[cekim]

This is linux under low load with speed-step and c-states disabled in the bios (clocks are 4.7/4.2):
more /proc/cpuinfo | grep -i mhz
cpu MHz : 1241.250
cpu MHz : 4444.804
cpu MHz : 1243.945
cpu MHz : 2892.890
cpu MHz : 2930.273
cpu MHz : 2614.687
cpu MHz : 4248.750
cpu MHz : 3172.851
cpu MHz : 2177.343
cpu MHz : 2931.914
cpu MHz : 2780.273
cpu MHz : 2014.570
cpu MHz : 2722.968
cpu MHz : 4262.109
cpu MHz : 1200.820
cpu MHz : 2528.671

If I peg it to 100% load, they all go to to ~4700. It behaves like I have a "performance" speed-step governor running (except more stable so far at any given clock/voltage).

In fact, it is even "twitchier" than the performance governor. It blips to max very quickly compared to the performance governor and lightning fast compared to the power-save governor.

Here we are under full load:
more /proc/cpuinfo | grep -i mhz
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4699.921
cpu MHz : 4700.390
cpu MHz : 4699.921
cpu MHz : 4700.039
cpu MHz : 4699.921
cpu MHz : 4700.156
cpu MHz : 4699.921

[cekim]

Another 16 hrs now (following 26 hours of stable), but this time with speed-step enabled in the BIOS. No change in behavior within linux with or without. I was mistaken, it is just as twitchy (prone to jump to max freq) with and without speed-step enabled in linux if the performance governor is enabled.

BUT, I did not re-enable C-states.... (previously auto when I was having random shutdowns).

Commanded freq/voltage:
4.5/4.2 @ 1.2v/1.0 (adaptive CPU at 1.2 turbo, offset cache at +.200).

This sort of setup was not stable with C-states at "auto". It would run fine under load and crash when randomly while idle, but no chance of making it to 16+ hours as it has 2x now (at 4.7@1.25 and now 4.5 @ 1.2). Going to let it run until tomorrow like this before I try the next experiment with C-states.

[cekim]

Update:
I got impatient after 24 hours of stable at 4.5/4.2 (see above). I've enabled speed-step AND C-states, but disabled the "enhanced idle" option which smelled suspcious given that my issues relate to what the chip does when idle not under load and more typically it crashes in the middle of the night doing exactly nothing. So far it is looking like I am on to something with that. Very different and much better behavior.

So, I:
1. re-ran Aida for an hour at 4.7/4.2 1.25/1.0 (adaptive/offset respectively) - no issues
2. went back into linux and did my usual exemplar stresses and then idled:
a. my desired CPU frequency behavior is now back (it drops down to 1200 and stays there with typical "surfing the web" loads and spins up quickly to 4700 with real load.
b. my idle temps are back down where they should be (25-28C core temp vs 30-32C)

So far, so good after 8 hours. This is what idle _should_ look like:
cpu MHz : 1242.890
cpu MHz : 1200.117
cpu MHz : 1199.882
cpu MHz : 1200.117
cpu MHz : 1200.000
cpu MHz : 1200.000
cpu MHz : 1200.000
cpu MHz : 1200.000
cpu MHz : 1208.437
cpu MHz : 1219.453
cpu MHz : 1282.382
cpu MHz : 1202.109
cpu MHz : 1200.703
cpu MHz : 1201.875
cpu MHz : 1201.289
cpu MHz : 1256.718

I haven't explored higher cache frequencies voltages much other than to see that:
a. I don't gain significant performance from 4.2->4.5 (far less than 1:1 on already small percentage increase)
b. so far I appear to need significant cache voltage to do that (1.3 vs 1.0 for 1.2)
c. still hunting for rock solid stable with cache above 4.2 w/low voltage
d. yet more variables, so I will explore that once I am convinced the idle issue is gone.

[meankeys]

[QUOTE=cekim;5

I haven't explored higher cache frequencies voltages much other than to see that:
a. I don't gain significant performance from 4.2->4.5 (far less than 1:1 on already small percentage increase)
b. so far I appear to need significant cache voltage to do that (1.3 vs 1.0 for 1.2)
c. still hunting for rock solid stable with cache above 4.2 w/low voltage
d. yet more variables, so I will explore that once I am convinced the idle issue is gone.


Did you notice higher temps with higher cache voltage?
higher cache speed over 37 i need a voltage adjustment to say 1.357v that will pull 30+ amps on the 12v rail and the percentage gain vers performance gain just isn't there for me. still testing.

[cekim]

meankeys wrote:
Did you notice higher temps with higher cache voltage?
higher cache speed over 37 i need a voltage adjustment to say 1.357v that will pull 30+ amps on the 12v rail and the percentage gain vers performance gain just isn't there for me. still testing.

Definitely higher temps.

With the BW-E, I'm stuck at 3.7 for stable @ 1.29v. It can do 3.8 at 1.32, but if I run memory stress long enough, I get single bit corruption reported in stress-app. Despite no crash of BSOD, if stressapp sees something, it will _eventually_ bite you somewhere. So, not worth it.

With HW-E, that limit, for my chips, remains 4.2.

Between 3.0 and 4.2, I see linear improvement in my various use-cases. 10% more cache gets me 10% more performance. So, 3.7->4.2 is another ~10% that I can't get, but the chip will do what the chip will do.

[meankeys]

Nice work guys
some very interesting info. thanks for posting.

Even on SS running RB over 1.4v will all most always crash on me. when running the encoder part of the bench it will pull down the SS to -18c = 66c core temp. it will crash every time.

speed step and c states off also multi core enhancements off

[meankeys]

Yes HW-E I do much better Core to Cache - performance.
BW-E not so good. I read you mention the 6950X has trouble @ 5 GHz your right. I also have a 6900X that seems to be a real dog past 4.8 GHz that is in my RVE. I am going to move it to my X99 DELUXE II and see if it will perform better.