For unlocked procs, what provides better performance – high cpu multi or high bclock?

[SteveRo]

Voltage settings - for the 1st pass through the tests I was thinking of keeping v's constant, this takes one variable out of the equation and also makes a 1st pass quicker to execute (no time fidling to find lowest v's). Depending on the results of the 1st pass we can then try to find lowest v stability and test again.

i am planning on using intel max v's for all but cpupll and dramv - probably much more than is needed but that's ok for the 1st pass - so planning to use the following -
cpuv-1.4v
qpivtt-1.4v
cpupll-1.8v
dramv-1.66v
LLC - default
I think I can leave everything else on default

Anyone have any thoughts on this?

(edit - i am not worried about cooling - will use chilled water)

[Ghostleader]

So for 980x mutli testing we have the following -

21x190 (3990) mem and qpi something like 190x8=1520, 7-7-7 1T, qpi at 16
30x133 (3990) mem and qpi something like 133x12=1596, 7-7-7, 1T, qpi at 24
35x114 (3990) 14x mem for 1596 and qpi at 28
19x210 (3990) mem and qpi something like 210x8=1680, 7-7-7 1T, qpi at 16

Interesting, but what about UnCore Frequency or is it that you refer to when you say qpi.

I bet my that Bclk 210 wins this one, because it has highest Bclk, ramspeed and uncorespeed

[famish99]

So I can't really speak from experience concerning unlocked Intel chips, but I've had unlocked AMD processors in the past and also I work for a semiconductor company doing silicon validation so we obviously get unlocked processors (engineering samples are the best... but sometimes you get the occasional dog).

In my experience, you want a combination of both. As many others have pointed out, core frequency is the biggest determinant of speed for these chips (mostly cause the cache architecture is so efficient) so keeping the rest of the platform within spec tends to make overclocking a bit more effective as it keeps the heat production lower and allows you more headroom for the core. That is to say that overclocking ram/uncore doesn't yield fruit, but within reason. However, increased bclk is also necessary because there becomes a point where higher multipliers are detrimental based on how the PLL is designed. As you increase multipliers, the clock jitter is amplified by the higher multiplier and can cause lower max frequencies.

[SteveRo]

Interesting, but what about UnCore Frequency or is it that you refer to when you say qpi.

I bet my that Bclk 210 wins this one, because it has highest Bclk, ramspeed and uncorespeed

good catch - I will fix that in my previous posts - meant to say uncore not qpi

[SteveRo]

So I can't really speak from experience concerning unlocked Intel chips, but I've had unlocked AMD processors in the past and also I work for a semiconductor company doing silicon validation so we obviously get unlocked processors (engineering samples are the best... but sometimes you get the occasional dog).

In my experience, you want a combination of both. As many others have pointed out, core frequency is the biggest determinant of speed for these chips (mostly cause the cache architecture is so efficient) so keeping the rest of the platform within spec tends to make overclocking a bit more effective as it keeps the heat production lower and allows you more headroom for the core. That is to say that overclocking ram/uncore doesn't yield fruit, but within reason. However, increased bclk is also necessary because there becomes a point where higher multipliers are detrimental based on how the PLL is designed. As you increase multipliers, the clock jitter is amplified by the higher multiplier and can cause lower max frequencies.

Does increasing cpupll volts reduce (or reduce the impact of) clock jitter?
How about uping the cpu clock drive voltage? - i usally run this at 900mv.

[famish99]

Does increasing cpupll volts reduce (or reduce the impact of) clock jitter?
How about uping the cpu clock drive voltage? - i usally run this at 900mv.

Increasing VDD_PLL will increase stability by reducing the amount of clock jitter, but this is hard to quantify and can add a good amount of heat as well. Lastly, PLL's are very sensitive devices (read break easily).

Increasing clock drive voltage increases overclockability because it increases clock slew rate and reduce setup time violations but it doesn't do much in the way of reducing jitter. But this too is hard to quantify.

Both of the optimal settings for these adjustments have been found via trial and error because short of that, you can only determine the optimal settings to these via circuit simulation.

[Johnny87au]

I'm not sure how low a bclock will boot but how about -
35x114=3990, 14x mem for 1596 and uncor at 28?

edit - what test should we use - I like pcmv for all around performance, should we do 3Dmv for the gamers in the crowd? What else?

Sounds good, Atleast it will final settle the truth as to what is better, higher multi or blck , Tbh im going with multi but lets see

[Johnny87au]

good catch - I will fix that in my previous posts - meant to say uncore not qpi

Can u run Super PI scores? I dont think their will be much a difference since the ram speed is only a tad faster.. super PI usually goes on cpu speed..

[SteveRo]

Can u run Super PI scores? I dont think their will be much a difference since the ram speed is only a tad faster.. super PI usually goes on cpu speed..

Looking at hwbot - looks like SuperPi 1M and wPrime 32m are most popular these days - so against each test case I think I will run

SuperPi 1M
wPrime 32m
3DMark Vantage
PCMark Vantage

Any thoughts?

[Johnny87au]

definately, gogogogo get sum results manggg

[shadewither]

Increasing VDD_PLL will increase stability by reducing the amount of clock jitter, but this is hard to quantify and can add a good amount of heat as well. Lastly, PLL's are very sensitive devices (read break easily).

Increasing clock drive voltage increases overclockability because it increases clock slew rate and reduce setup time violations but it doesn't do much in the way of reducing jitter. But this too is hard to quantify.

Both of the optimal settings for these adjustments have been found via trial and error because short of that, you can only determine the optimal settings to these via circuit simulation.

I have tried vpll on a Xeon W3565 on its default VID 1.21875v around 4.1G (on R3E).
I have tried vpll 1.2/1.3/.../1.81(default)/1.85/1.9 on LinX "unstable" frequency 4.1G which usually BSOD during the second half of a LinX loop.
I found that:
vpll 1.3/1.81 => best, usually BSOD during the second half of a LinX loop
vpll 1.2/1.7/1.85 => worse, usually BSOD during the first half of a LinX loop
vpll 1.4/1.5/1.6/1.9 => worst, BSOD as soon as calculation begins (I mean, when memory allocated and all cores begin to load). And the BSOD code is different.

And for higher voltages/frequency, e.g. 1.4v/4.5G, vpll 1.3 seems better than vpll 1.81.

Any explanation on this? Or merely a corner case?

[SteveRo]

@Mr. shadewither - a lot of folks are finding lower cpupll works well for oc's and some claim this results in lower temps also.
default on the ggbt x58UD7 is 1.8v, intel max is 2v, I seem to be able to get good 4.7+ oc's with 1.5v, haven't tested at lower than 1.5v.

[cory1234]

Both of these settings passed 50 IBT runs. Increasing my QPI decreases my GFLOPs. Does anyone know why my i5 750 was getting more GFLOPS than my new i7 875k?

[famish99]

Yeah it seems like Intel spec the VDD_PLL too high and many of the stories on the forum point to that, I haven't had the opportunity to confirm this myself because my board won't go below 1.8v on VDD_PLL. It's been mixed results from what I've experienced at work as well, but sometimes it helps. It just seems like in the case of the Nehalems that lower is better.

The only thing that I find odd about this experiment you're going to run is that you're using the same core speed. In the benches I've run, given the same core speed, a faster qpi/uncore will increase performance. I think a more interesting experiment would have been to see what allows for high overclocks, using a high multiplier and moderate blck or a high bclock and moderate multipliers

[SteveRo]

@famish99 - yes, same core speed is phase 1.
Phase 2 is to find the best oc for a given cpu multi and compare again - i'm thinking of using 20xlinx to determine stability

[famish99]

Phase 2 is to find the best oc for a given cpu multi and compare again - i'm thinking of using 20xlinx to determine stability

I look forward to the results of this experiment then, although I think it might align with what I've seen at work.

[some_one]

Here's some results using a 655k. 100MHz BCLK with 46x multi and 200MHz BCLK with 23x multi. Ram speed and timing the same (12x100 and 6x200) only differences AFAIK were

1. Vtt 1.1V for 100MHz and 1.24V for 200MHz
2. Uncore, 18x100MHz and 18x200MHz.

The increase in uncore helped a little, maybe through L3 Cache.

[Ghostleader]

I ran a comparison between Bclk 133 and 200 at 4,0GHz (4C/8T) with y-cruncher (BenchmarK Pi, Multi threads, 1 billion digits test).

It was run with my i7 975, RIIE, 6Gb of ram.

Bclk 133.4
Core ~4002
UnCore ~3200 (x24)
QPI ~3200 (x24)
Ram 800,5
Timing 777 21 1T

vs

Bclk 200,0
Core ~4000
UnCore ~3200 (x16)
QPI ~3600 (x18)
Ram 800
Timing 777 21 1T


Result

Well the screenshot talk for it self.



Not much of a different but there´s one and it favor the 200 Bclk.

Any thoughts anyone?


EDIT: I ran three of each and took the average and the trend is the same as in the screenshot, the screenshot is from the last run of each.

Before each run I reboot the system.

The variation between run-to-run was less then 1%.


I ran it with W7 64bit in diagnostic mode so that I got the least amount of interference from servies or other things.

[shadewither]

@Ghostleader
Thanks for your comparison.
Higher blck for higher performance is expected.

However, how many times have you run y-cruncher?
I thought that y-cruncher has a ~1% run-to-run variation.

[famish99]

The y-cruncher results speak for themselves. It's somewhat expected that at a give core frequency that higher platform frequencies will yield better results as the execution units are spending less time waiting for memory accesses. I'm just wondering what the results are for the phase 2 experiment.

[Ghostleader]

@Ghostleader
Thanks for your comparison.
Higher blck for higher performance is expected.

However, how many times have you run y-cruncher?
I thought that y-cruncher has a ~1% run-to-run variation.

I ran three of each and took the average and the trend is the same as in the screenshot, the screenshot is from the last run of each.

Before each run I reboot the system.

The variation between run-to-run was less then 1%.


I ran it with W7 64bit in diagnostic mode so that I got the least amount of interference from servies or other things


I notice a small different in Round Trip Latency, it´s 52 for Bclk 133 and 51 for the Bclk 200, I don´t know from which channel it reports it from but the change is that it´s the same trend from each channel.

So one lower for Bclk 200 probably also contribute to a lower y-cruncher result.

[SteveRo]

I ran three of each and took the average and the trend is the same as in the screenshot, the screenshot is from the last run of each.

Before each run I reboot the system.

The variation between run-to-run was less then 1%.


I ran it with W7 64bit in diagnostic mode so that I got the least amount of interference from servies or other things


I notice a small different in Round Trip Latency, it´s 52 for Bclk 133 and 51 for the Bclk 200, I don´t know from which channel it reports it from but the change is that it´s the same trend from each channel.

So one lower for Bclk 200 probably also contribute to a lower y-cruncher result.

Good morning,

Thanks for the data, so looks like small but real advantage for high bclock over high multi.

I am running my phase 1 now, planning to do 3 runs of all tests at each multi, I should be able to confirm what you have captured.

[Ghostleader]

@SteveRo

You´re welcome

I really looking forward to see your results of the phase 1 tests.


I think that if one can maintain the same core, uncore,ram frequency and ram timing at different Bclk frequency the outcome going to be really close.

[Brama]

I ran a comparison between Bclk 133 and 200 at 4,0GHz (4C/8T) with y-cruncher (BenchmarK Pi, Multi threads, 1 billion digits test).

It was run with my i7 975, RIIE, 6Gb of ram.

Bclk 133.4
Core ~4002
UnCore ~3200 (x24)
QPI ~3200 (x24)
Ram 800,5
Timing 777 21 1T

vs

Bclk 200,0
Core ~4000
UnCore ~3200 (x16)
QPI ~3600 (x18)
Ram 800
Timing 777 21 1T




Not much of a different but there´s one and it favor the 200 Bclk.

Any thoughts anyone?

Higher QPI could explain the difference.

[radaja]

Both of these settings passed 50 IBT runs. Increasing my QPI decreases my GFLOPs. Does anyone know why my i5 750 was getting more GFLOPS than my new i7 875k?

well what gflops did you 750 get?both pics you posted are the 875K with different qpi speeds.