Kaby Lake / Asrock disappointment, RAM weirdness

[Madpoo]

Originally Posted by Prime95

Weirdness: Why does setting ram to DDR4-2400 or DDR4-2266 give worse performance than a DDR4-2133 setting??

Does it change the latency at all? At some point maybe the faster speed provides incremental gains but increased latency takes it back, with interest?

[retina]

Originally Posted by Madpoo

Does it change the latency at all? At some point maybe the faster speed provides incremental gains but increased latency takes it back, with interest?

Yes, perhaps. The internal mechanisms of DRAM storage haven't really got any faster for many years. Only the interface logic has increased the clock speeds and the bus widths have got wider. The actual retrieval times from initial addressing to final output haven't really changed with any significance, and can be (marginally) worse if the clock speed and retrieval times are not exact multiples of each other.

However, all of the above can become moot if the software is able to prefetch data before usage and hide any latency. So it depends upon what you are doing as to how it affects the final application.

[bloodIce]

I see a moderate improvement with Kaby Lake, but my LL results started to suffer from ROUNDOFF errors. I am using mprime 28.10 on Ubuntu 16.10. The FMA3 FFT length is 5760K and I did not have a single problem with my previous processor (Haswell). I am a bit disappointed after upgrading, because now I cannot do a (reliable) LL above M10,000,000. The whole point of my upgrade was a faster LL. Is it the code? I have Corsair RAM certified to work on 3200, but currently working at 3000, so it should not be that.

[kladner]

[QUOTE]I see a moderate improvement with Kaby Lake, but my LL results started to suffer from ROUNDOFF errors.[/QUOTE]What are the specs for the new CPU, particularly core clock and voltage?
Please excuse the unnecessary question. My mind was in another thread.

[Prime95]

Originally Posted by bloodIce

my LL results started to suffer from ROUNDOFF errors... I have Corsair RAM certified to work on 3200, but currently working at 3000, so it should not be that.

Try bumping your RAM voltage or CPU voltage. My Skylakes are all happy undervolted by 30 or 40 mV. My KabyLake is only happy at -20mV. My Skylakes ran the RAM at 1.25V, KabyLake required 1.30V.

It is a pain, but trial and error should eventually find the settings you need for stability. Do double-check assignments for several weeks. So far I have 1 week of error-free operation - another week or two and I'll call these settings stable. If you cannot get the CPU stable at DDR4-2400 then RMA the chip.

[bloodIce]

Thank you for the advice. I just want to say that DDR4@2400 is not a problem, however obviously my productivity/sek will decrease (at the end this is not only a borg). You are right that I am at 1.2V (stock), but this is because I have 64GB and higher voltage at 3000MHz creates a heat dissipation issue. However, I will try 1.25V at the RAM with slight downvolting my CPU (i5-7600). May be I was not correct when I said "moderate improvement" over Haswell, there is nearly doubling, but come on, my RAM is at >2x in comparison to DDR3.
I have one more first time LL to finish in a day or two, and I am sorry but I will submit it. Otherwise, after that I will do DC until it is safe to go for LL.
The thing is that ECM and Pm1 work fine and are more productive.

[lycorn]

Why do you say ECM works fine? Have you found any factors yet?
That said, it would be nice to have you doing GMP-ECM on a 64 GB RAM machine... Some monster factors might be popping up.