Why does increasing memory speed increase latency?

[bwanaaa]

The only thing I changed is speed from 3600 to 3800.

 

[EarthDog]

Because you went from 3600 to 3800 and the Infinity Fabric/FCLK ratio broke 1:1.

 

[bwanaaa]

Because you went from 3600 to 3800 and the Infinity Fabric/FCLK ratio broke 1:1.

According to zentimings I did not. FCLK, UCLK, and MCLK are all 1900

 

[EarthDog]

I see now... no idea offhand why that would increase as your timings did not.

 

[bwanaaa]

according this fellow on you tube, raising the memory speed sucks more power.

This has adverse consequences on the cpu . So even though I am looking at latency, this is not a strict memory benchmark. The cpu has to initiate those read, write and copy instructions.

 

[mackerel]

Did you run them more than once to check the results? Maybe there was a bad run.

 

[bmwbaxter]

Did you run them more than once to check the results? Maybe there was a bad run.

+1

Also were both runs done the exact same way? something else wasn't also being done in the back ground?

 

[bwanaaa]

EXACT same way? These days there are so many processes in the background it's hard to know. Yea, Afterburner, RTSS, and who knows what else. I'll rerun them and make sure all non essential processes are terminated first.

 

[bwanaaa]

ok.
I set the pc to safe mode.
Cleared CMOS
Set vDIMM to 1.46v for all testing
CPU settings were stock for the first tests and then PBO was enabled for a few more.
The data were gathered using Aida and tabulated in Google sheets here:

https://docs.google.com/spreadsheets/d/1PhzFyWciQLrQR93lxPpDh--641xLisUfm0CnWRcxSpc/edit?usp=sharing

best results were at 4000/2000. Although I could run at 4066/2033, the results were worse.
Also, PBO causes L3 cache read speeds to take a massive hit.

I will be repeating some of these tests in regular mode. I will stop certain processes and services before each run.

 

[Woomack]

When you set a higher memory clock then all settings under auto, adjust to the "table values" which you can't see. It affects memory timings and some other settings. Another thing is when you think it's stable but it really isn't then errors are being corrected and calculations are losing cycles. Then you get worse results too.

 

[Zerileous]

Yep offhand I noticed tRC and tRFC both got set higher by the motherboard. Not sure if its enough to account for the difference, but they do have an effect.

 

[Woomack]

tRFC affects latency more than CL ... I mean going down with tRFC by 200-300 usually gives better results than with CL from 16 to 14

 

[bwanaaa]

Thank you all for your help. I got the latency fixed by tightening those secondary timings. Now I'm here: