How to get better latency

[Ausknight]

Most people cant notice the subtle differences in how a PC runs... if your a bencher / gamer / OC'er, you can notice even the slightest problems in a system.

Im still working on this Ram to try get it at 1600.. its at home P95'ing now so I will know if its stable in another 2 hrs. i think its at 1600 7-7-7-24 or 8-8-8-24.. something like that, i dont have my worklog with me atm.

I would love to be running 1866 or 1900 but my Stupid N790i Ultra SLI causes too much Vdroop and it gets scary with voltages on the CPU.

[GSKILL TECH]

ehume, sorry I'm a little late. The 860 processor only likes the 12x memory multiplier. So your best results would be BCLK 167 for DDR3-2004. I noticed you did try it and it froze, what memory timings and voltage did you have it set at?

The reason why you are only allowed certain memory frequencies is because BCLK 133 with memory multi 6, 8, 10, 12 can only produce those values. For values in between you'll need to adjust BCLK accordingly.

Lowering frequency and tightening timings does compensate for some speed, but in P55 tests the higher frequency does make quite a difference. We tried all sorts of settings to get the same results with lower frequency and timings, but especially DDR3-2000+ it is another level that cannot be matched. For maximum speed, you'll definitely want to attain the highest memory frequency. You have experienced that as well from 1333 to 1600, so you can imagine what speeds will look like at DDR3-2000+.

Thank you
GSKILL SUPPORT

[ehume]

I think you are absolutely right.

[ehume]

Bumping up against limits

My previous message was for RamJam's post.

I have decided not to push the BCLK, because I want my $290 CPU - the most expensive item in the machine I built - to last a long time. So a multiplier of 12 is my max with a BCLK of 133MHz. I know I am missing out on a scorcher of a machine, but until I get a SSD for a boot disk, I can't really use that performance.

I am impressed that Gigabyte's MB supports the 1600 DDR mode, even though it does not claim to. I am also impressed that the GSKILL memory supports this choice, even though it does not show up on the SPD, at least on CPU-Z.

I noticed that at DDR2000, 9-9-9-x works fine. 8-8-8-x would not get past POST. at DDR1600, 7-7-7-x works fine. 6-7-7-x fails POST. At DDR1333, 6-6-6-x works fine. 5-6-6-x fails POST.

A DDR of 2000 means a clock rate of 1000. So each cycle is 1 ns. What works for CAS latencies:

1000MHz 9 x 1.00 = 9.00ns
800MHz 7 x 1.25 = 8.75ns
667MHz 6 x 1.50 = 9.00ns
533MHz 6 x 1.875 = 11.25ns

Unfortunately, my MB will not accept an input of 5 for CAS latency, so I could not test 533MHz 1.875 x 5 = 9.375ns. That would have been interesting.

What I think I have established is that the Gskill F3-16000CL9D will allow a CAS latency down to 8.75ns. Somewhere between that and 8.0ns (that failed post at DDR2000) is the rock bottom value that will work.

What I would like to see is for GSkill to publish the minimum latencies in nanoseconds that the memory will support - for each of the many settings. Then we can calculate our own cycle settings, based on the current clock rate (the clock rates do vary a bit).

How about it, guys?

[ehume]

I especially would like to know the lower limits of tRAS, tRFC and tRC.

[ehume]

And yet With this post the guy was using the same part number but getting down to CAS latency of 7. Of course, this blows my entire theory into a cocked hat.

So, GSkill tech guys, what goes? What are the limits and how do we calculate them?

[ehume]

So I took it as a challenge. I set my BCLK to 166, not 167. Turned off the turbo boost - I forgot to do that last time, so that might be why the system quit. I set the timings to 9-9-9-28 with a tRC of 47, a tRFC of 88. RAM voltage is set to 1.66, which results to 1.648V.

Sandra: bandwidth 23.870, latency 60.2, speed factor 52.3.

You're right. That is a big kick in the pants. I'll just have to get a better CPU cooler if I decide to work at this level.

[ehume]

Running the WE:

Processor = 7.6 (up from 7.5)
Memory = 7.7 (up from 7.5)
Graphics & Gaming Graphics = 6.9 (unchanged)

And the tempt ran up to 59C briefly. I definitely need a stronger cooler for playing at this level.

[GSKILL TECH]

If you simply keep the CPU frequency at stock level you should not have an issue. 167 x 17 should give you near stock frequency and temperatures.

Thank you
GSKILL SUPPORT

[Ausknight]

Hello,

I have these chips and I cannot get them stable for the life of me,
Instead of the 2000fsb I want to run them at 1800,
I currently have them set to 9-9-9-27 1.65v as per specifications but they are not stable.

9-9-9-27 1.65v, I have all advanced times set to Auto. Should I really need to tighten times for 1800?

CPU can pass 20loop regular and high stress in Iburntest. and runs 8hrs P95, Smallfft (not much ram tested), as soon as I put it on blend it will fail after about 5mins. (rounding errors)

Some other info I have: Q9550, N790i Ultra-SLI, Antec1000 PSU, GTX295 and 8800GTS,

Does anyone have some recomendations to try for 1800fsb?

[ehume]

I tried the BCLK = 167MHz with the CPU multiplier at 17. It worked well. I am now pondering whether to use it with Turbo Boost.

What BCLK and RAM multiplier do you use to get DDR1800? 150MHz x 12?

[GSKILL TECH]

Yep, 19 X CPU multi to give you 2850Mhz.

Ausknight,

work on your voltages, 1800 FSB will require spot on voltages to run properly.

Thank you
GSKILL SUPPORT

[Ausknight]

Thanks for that, I have dropped it to 1.6v and its 3hrs P95 stable so far. once it gets at least 8hrs i will be happy.