Help me squeeze last bit of juice out of Kingston HyperX Predator 1866

[FlayedOne]

Hi,
Recently I became a lucky owner of 2x4 Gbs of Kingston HyperX Predator 1866 memory sticks, and since my rig uses integrated graphics of A10-5800K RAM performance is really important for my rigs graphics performance.

First thing I did(after overclocking NB to 2400) was setting them to [email protected] and stock timings for the Predator 2400: 11-13-13-30. It went great - the memory was perfectly stable. I tried tightening the timings and got them down to 11-12-12.

I don't know what to do with the 4th timing though. Right now it's set at 25(11+12+2), but the memory seems stable(passed 1 round of memtest86) even with a setting of 22(possibly lower). I didn't see any benchmark improvments associated with that setting though. Should I try and get this as low as possible?

And what should I do with the advanced timings? I can't set the command rate to 1(generates lots of errors), but the rest right now is set according to XMP for 1600, and the memory is perfectly stable. Still, it seems that after I lowered the other advanced timings from XMP values for 1866 reading from the memory actually got slower while writing got faster. WTH is going on? How should I set them to get the best results? Unfortunately they are totally ignored in most of overclocking guides and reviews...

Thanks in advance.

 

[redduc900]

You'd probably get better bandwidth results by lowering the effective DRAM frequency, and in turn tightening the timings back to their rated 9-11-9, including setting the Command Rate to 1T / 1N.

 

[Woomack]

It looks like these 1866 Predators are Hynix based ( like 2400 that I was testing ).
Try:
1866 8-10-9-27 1.65V
1866 8-11-9-27, 1.65V
2133 9-11-10-28 1.65V
2400 11-12-11-32 1.65V
2600 11-13-12-34 1.65-1.75V ( that one depends from IMC and may require higher CPU-NB voltage )

AMD don't really like Command Rate 1N and it probably won't run stable.
4th timing for hynix is usually like CL+tRCD+tRP=tRAS-2 , but can be stable lower or higher.

Like redduc900 already said, I would check performance on lower clocks with tighter timings as memory bandwidth for AMD is limited by IMC so setting higher clock won't give any special difference but tightening timings will help in access times.
Usually best performance can be achieved by setting 1866-2133 memory clock and as high as you can CPU-NB clock ( which is limiting factor for AMD ).

 

[FlayedOne]

Thanks for the answers guys.

I tried getting tighter timings@2133 and here is what I got:
9-10-10-28 with command rate 2
10-10-10 28 with command rate 1

Both resulted in slightly worse transfers than 11-12-12-27 and a little better latency. After testing in 3DMark it seems that the APU agrees better with 2400Mhz though.

I was sure I tested it before, but you're right Woomack - 11-12-11 is stable! Thanks.

It seems I get the best results with 11-12-11-30@2400. I only tested 32,30 and 27 tRas though, so that may change (31 and 29 are likely to be even better)

The NB seems to be the weak link of the APU architecture - it's voltage is required to oc both the NB clock and the GPU, so if one wants to oc both he needs to settle for lower clocks.

Once again, thank you guys.