Timing rules

[Aquinus]

Leave HT frequency alone! It does nothing!!!

I found with Bulldozer and later CPU's anything over 2400mhz NB requires a lot more voltage if it's going to be stable at all.

Just to expand upon this, HT does nothing for CPU performance, I hardly say that it does nothing. For everyone else (who may not know,) the hyper-transport link is for all PCI-E and southbridge communication. Pushing HT too high makes your NB on the motherboard work harder (not just the CPU,) and won't result in increased performance. If for some reason you were ever able to saturate HT between video cards, SATA, and such, then you would see a performance boost but that's incredibly unlikely.

As for the CPUNB speed, I've found that with the 940, 960t, and 965 that 2.6Ghz was usually a sweet spot without needing too much voltage and offered the best memory performance at 3.4-3.8Ghz on the Phenom IIs I've owned in the past. Maybe I was lucky.

 

[d1nky]

i use all these higher frequencies for benching. im able to get loads better scores with cpunb around 2.7ghz HT about 2.7ghz cpu 5ghz+ ram from 2450-2550mhz

volts really depend on temps and i usually give about 0.1 extra on most.

i would thought my cpu being AMD copying an xmp profile is wrong as theres different stepping etc but id have to copy a 1800 Jedec which would be difficult to get stable at 5ghz cpu

 

[Aquinus]

i use all these higher frequencies for benching. im able to get loads better scores with cpunb around 2.7ghz HT about 2.7ghz cpu 5ghz ram from 2450-2550mhz

There also is a point of diminishing returns when a higher CPUNB doesn't offer you better performance. I've found that there is a sweet spot depending on what frequency you're running the core at, it's just a matter of finding it. In this case though, higher isn't always better.

 

[d1nky]

i have noticed this effect, its the same when my cpu goes above 5.1ghz it has a negative effect.

plus theres only a certain amount of data the cpu can process no matter the clock, so i think its kind of at its limits.


however i did boot with a 2.9ghz cpunb and ran a maxmem with stock speed and setting and well it was pretty darn good. definitely not stable tho.

 

[LAN_deRf_HA]

I'd like to see if anyone else can confirm this tweak. Lowering vccsa made a consist, however small, score improvement. Note this is on a platform where vccsa and vccio are supposedly linked. I say supposedly because according to aida vccio didn't get any lower which is just confusing.

vccsa 0.925v

vccsa 0.8v

 

[EarthDog]

You know, it's exceedingly difficult to find information on secondary timings, as I'm sure you're aware (I assumed you tried to Google it first). My understanding of DRAM timings in general is that many timings overlap with respect to when they actually occur during any memory operation and several timings occur within the primary timings. The real benefit from reducing secondary timings would really be to stabilize lower primary timings at any given frequency as opposed to actually improving performance. Beyond that, I'm not sure if I can be very helpful.

Best post in this thread so far...

I'd like to see if anyone else can confirm this tweak. Lowering vccsa made a consist, however small, score improvement.

Those differences are WELL within the margin of error/differences between runs...as you mentioned in the quoted post below, actually.

 

[drdeathx]

ive been doing this also

what ive learnt, two days studying.... got a headache as well lol

http://forum-en.msi.com/faq/article/ddr3-memory-timings-explained


in my notes i got some formulas for better timings.

so far im 2500mhz 11-13-13-31 1T 300ns 1.65v fully stable


heres some for rock solid stable:

TRAS = CL + TRCD + TRP (+/-1)

TRC = TRAS + TRP

TWR = TCL-1 + burst length/2 + TWTR


some for tighter:

TRC = CL + TRAS

TRAS = TCL + TRCD + TRP/2 + 2


What im looking for is if there is a calculation to guage whether higherspeed/looser timing vs lowerspeed/tighter timing

its easy doing a maxmem/aida test but theyre inconsistent and are based on cpu/cpunb/ht which changes with frequencies

I tested memory, sometimes 2133MHz at cas 9 gives better performance. Maybe you can get 2400MHz cas 9 then you have a winner

 

[LAN_deRf_HA]

Those differences are WELL within the margin of error/differences between runs...as you mentioned in the quoted post below, actually.

I'm not joking when I say consistent. I test and test and test and I can ONLY get those scores with those respective voltages. I'm very familiar with the score ranges.

 

[d1nky]

these timings ^^^ 11-13-13 allow me to get my speeds quite high. which in benchmarks is doing me well.

i need to tweak more to get my timings down. 2133mhz at CL11s is pretty mundane. im tied between speed or tight timings lol

 

[LAN_deRf_HA]

Finally figured it out. It was a "boot variance" coincidence. Idk what the proper term is but some benchers may have noticed scores are better on some reboots more than others regardless of how long you let the system settle. So I left the voltage alone, rebooted, ran 10 times looking at the peak scores. Rebooted 6 more times repeating that sequence and found the scores to be up and down almost every other time which explains why I thought it was coinciding with the voltage change since I kept changing it every other time.

 

[d1nky]

LOL kind of cat and mouse effect!