Clarkdale, SS and PI32M

[massman]

Sometimes a man needs a break from all the high-end top-notch hardware overclocking and go back to basics. For me, basics means Super PI 32M with some challenges; in this case: Clarkdale and single stage cooling.

First, I looked up all the Clarkdale SuperPI 32M Single Stage scores in the HWBOT database, which rather surpringly resulted in a list of a mere 30 results. Then I ordered that list according to PI time and give it some nice Vbulletin tags.

431.672 seconds - pro with Core i5 655k @ 5388.00 MHz ( 898 CL6.0 6-5-17 )
439.047 seconds - 0Ro! with Core i5 661 @ 5589.60 MHz ( 972 CL8.0 9-8-24 1T )
443.000 seconds - Ross with Core i5 670 @ 5315.00 MHz ( 924 \N )
444.343 seconds - momoto with Core i5 670 @ 5586.50 MHz ( 798 CL6.0 6-6-19 1T )
450.547 seconds - RULE with Core i5 670 @ 5670.00 MHz ( 840 CL7.0 7-7-20 )
450.812 seconds - rambo R&B Team with Core i5 661 @ 5400.00 MHz ( 0 \N )
454.422 seconds - Chuchnit with Core i5 670 @ 5336.00 MHz ( 928 CL7.0 8-7-21 1T )
456.781 seconds - Brolloks with Core i5 670 @ 5400.00 MHz ( 0 \N )
458.906 seconds - RULE with Core i5 670 @ 5593.10 MHz ( 932 CL7.0 7-7-20 )
462.141 seconds - Vanz with Pentium G6950 @ 5272.00 MHz ( 0 \N )
467.437 seconds - Unseen with Core i3 530 @ 5192.00 MHz ( 0 \N )
472.703 seconds - majkel with Core i3 550 @ 5184.50 MHz ( 864 CL6.0 7-6-18 1T )
475.062 seconds - Brolloks with Core i5 661 @ 5200.00 MHz ( 0 \N )
477.250 seconds - kaidtor with Core i3 540 @ 5221.00 MHz ( 908 CL8.0 8-8-24 )
477.563 seconds - NAMEGT with Core i3 530 @ 5280.60 MHz ( 720.1 CL7.0 7-6-20 1T )
479.156 seconds - ScunnyUK with Core i3 530 @ 5170.00 MHz ( 705 \N )
479.438 seconds - isulk with Core i5 661 @ 5225.00 MHz ( 836 \N )
482.344 seconds - NAMEGT with Core i3 530 @ 5280.20 MHz ( 720.1 CL7.0 7-6-20 1T )
486.678 seconds - [GF]Duke with Core i5 670 @ 5359.80 MHz ( 794 CL7.0 7-7-20 )
490.792 seconds - Massman with Core i5 661 @ 5040.00 MHz ( 0 \N )
492.134 seconds - kaidtor with Core i5 660 @ 5304.00 MHz ( 816 CL8.0 8-8-24 )
494.125 seconds - muzwa kai with Core i5 670 @ 5400.00 MHz ( 900 \N )
494.271 seconds - kaidtor with Core i3 530 @ 5170.00 MHz ( 705 CL8.0 8-8-24 )
497.078 seconds - muzwa kai with Core i5 650 @ 5216.00 MHz ( 1043 \N )
499.092 seconds - systemviper with Core i3 560 @ 5012.60 MHz ( 1002.5 CL7.0 8-7-20 )
503.656 seconds - majkel with Core i3 540 @ 5014.50 MHz ( 872 CL6.0 6-6-18 1T )
503.828 seconds - koziro with Core i5 655k @ 5208.10 MHz ( 840 CL8.0 8-8-21 1T )
505.453 seconds - Rocx with Core i3 530 @ 5126.60 MHz ( 932.1 CL10.0 10-10-30 )
517.250 seconds - ocgmj with Core i5 661 @ 5005.30 MHz ( 770 CL9.0 9-9-24 )
534.282 seconds - johnksss with Core i5 661 @ 4920.00 MHz ( 600 CL7.0 7-7-16 )

What was I thinking?! Of course, when I need to find the most efficient (time vs cooling) Pi32M scores I need to check out Youngpro's profile first! That being said, I'm also impressed by Ross' run. I'm sure he would get really close to Youngpro with a bit higher CPU frequency.

So, after playing around a bit with this H55 board focussing on the Intel IGP overclockability, I went for a round of PI. Of course I don't want to use the 2:8 divider (which is totally on Clarkdale), so I needed to use the 2:10. It's also known that Clarkdale is pretty bad in memory overclocking, so anything south of DDR3-2100 I would be extremely pleased with.

So, without touching the subtimings or doing a copy-waza I reached this score:


(clickable)

[hwbot=1078597]submission[/hwbot]

Now, I need to go back to my lab and try to beat the other scores

[[XC] Oj101]

My word James' efficiency: 2,325,848.736
Your efficiency: 2,340,554.88

You're almost there!

[G.Foyle]

Pretty close, but that yard left is the longest

[0Ro!]

ugh, looks that Im very bad in efficienty (shame on me ) ....need more SPI studying to improove the result

[Calathea]

What about the 2:8 divider? I've only had one clarkdale, an i3 530, but on that cpu 2:6 were bad. It could make 235 bclk 2:6 but 250 2:8 (water).

[El Gappo]

Jame's superior accent tweak is clearly giving him an edge.

[youngpro]

wow dude your ram clocks are impressive, and very important of course with clarkdale or lynnfield,i think with tight subtimings, copywaza and if you can slightly subzero ram for 1080 7-7-7 then it will be much faster than my time mate

[Leeghoofd]

I want meee rams back :p You get the Flares instead , that's a real challenge then :p

Go go PJ !!

[massman]

What about the 2:8 divider? I've only had one clarkdale, an i3 530, but on that cpu 2:6 were bad. It could make 235 bclk 2:6 but 250 2:8 (water).

The 2:8 divider gives the WORST clock-per-clock performance on Clarkdale. You kind of have the proof already in your reply; 2:6 is much tighter and therefore doesn't allow that high BCLK. 2:8 is a lot less tight and allows high BCLK. Here's some numbers Chew* posted a while back:

- 8m 23.109 @ 2:8
- 8m 05.531 @ 2:10

That was on the ASUS Sabertooth, though ... the performance drop on that board is particularly huge. Might be an ASUS-thing, not sure. This GBT board has the performance bug to a bit smaller extend:

- 7m 36.500 @ 2:8
- 7m 34.938 @ 2:10

Since then, I'm staying away from the 2:8 divider just for sake of it. Also, memory clocking on Clarkdale is a bit of a challenge, so it's more fun.

Jame's superior accent tweak is clearly giving him an edge.

I think there's something magical about his clinically set-up bench room. Perhaps having the white table and white wall seemingly flow into each other helps shaving off those last few seconds.

wow dude your ram clocks are impressive, and very important of course with clarkdale or lynnfield,i think with tight subtimings, copywaza and if you can slightly subzero ram for 1080 7-7-7 then it will be much faster than my time mate

No, it's not impressive.

The ram clocks are okay. When I hit DDR3-2200 ... then it's good.

I want meee rams back :p You get the Flares instead , that's a real challenge then :p

Go go PJ !!

Gamer has my (supposed-to-be) Hyper!

[westsider]

What about the 2:8 divider? I've only had one clarkdale, an i3 530, but on that cpu 2:6 were bad. It could make 235 bclk 2:6 but 250 2:8 (water).

with clarkdale you should clock like: DRAM 2:10 and QPI 2:20(or above)
otherwise you get memory bandwidth that sucks!!!!
only 2:10 memory and 2:20+ QPI gives 12K+ memory Read

any other Setting gives about 8-9K Ram Read




most of the lil clarkys from i3 family could hit only 2050MHz
everything above 2070MHz wasnt possible
but my ram does 2700+ AIR

so its only y clarky thats fault in this case

[westsider]

and never disable these


example
http://s56.radikal.ru/i151/1009/ec/3fd63f105935.jpg
disabled

[massman]

My word James' efficiency: 2,325,848.736
Your efficiency: 2,340,554.88

You're almost there!

Efficiency-wise, I'm already there. Just can't get my phase to drop to -20°C with clarkdale. The read-out on the M3F is indicating -6°C

Small copy-waza. Also had a 23s flat, but pressed space bar and couldn't re-produce the run

[dinos22]

its more about the QPI link clocks than anything else....get that high and you're good to go

massman > GBT does not have a bug at all with 2:8. The bug is in your lack of attention in testing (check your screenshot again and you will see why there was a couple of seconds difference )

[massman]

Uh-oh, the Gigaussie PPolice is here!

As I mentioned already, I switched to a different mainboard to give this thing another spin. The M3F is giving me an extra 50 to 80MHz on the core most likely due to the tolerance for a bit more Vcore. In the end, CPU MHz does make the biggest difference. In terms of memory overclocking, the M3F doesn't do as well as the little H55 board, maxing out at DDR3-2110 so far. But, with the extra MHz on the core, I was able to push out semi-decent 'no-compromise' runs.

Here's my run with CW (3,5s gain over no CW)

[westsider]

its more about the QPI link clocks than anything else....get that high and you're good to go

massman > GBT does not have a bug at all with 2:8. The bug is in your lack of attention in testing (check your screenshot again and you will see why there was a couple of seconds difference )

only QPI doesnt bring anything
try this at home
give highest possible QPI and ram at X6 and you ll get poor bandwidth
ram X10 and QPi 22 for example and you get huge gain

[massman]

And another board on the table. This time the old-skool GD80 that I've had since, what, a year now? Obviously I cannot expect this board to pwn the new PCB designs out there, but it turned out to be quite a handful.

Unlike the two previous boards, this one will not clock memory that nicely. DDR3-2045 seems to be the limit for now and that already requires a 10-minute workaround to get it ready to run 32M . Will post the work-around if someone's interested.

Anyways, it does seem I can nail a bit higher clocks than on the other boards: 5.32G is consistently stable. Also had a nice 'competitive' (not really efficient) run going (7m24-ish), but it crashed after loop 24 () and the run after that rebooted at loop 21. FML. I'm sure tomorrow-night I'll produce a more competitive run in terms of end result. Efficiency-wise, unless I find the magic key to unlock DDR3-2100, it will be a no-go.

[dinos22]

only QPI doesnt bring anything
try this at home
give highest possible QPI and ram at X6 and you ll get poor bandwidth
ram X10 and QPi 22 for example and you get huge gain

^

Clarkdale's uncroe is fused. the design of clarkdale Uncore is a logic clock.
IMC controls ddr3 speed and performnace. and clarkdale put it into 45nm GPU die,
from CPU die L3 cache the path to GPU die is QPI bus.

Normally, QPI bus limits DDR bandwidth, in theory, when QPI bus is 6.4GT/s,
DDR3 speed 1333, the DDR3 banwidth will be 21GB/s ( in theory), also DMI bus is tru
QPI. thats why increase QPI bus affects DDR performance a lot. even you don't use
IGP. properly you might say with clarkdale the L3 smart cache path to IMC is the performance that we concern.

the platform is starved of bandwidth hence why QPI helps. QPI has a much stronger impact on pi times than timings but obviously I wasnt arguing about both simply that once you are at your optimum system settings look to max out QPI basically but not at expense of CPU clocks as massman pointed out. I've done some tests at 6Ghz with various QPI speeds and pushing QPI significantly improves efficiency.

I don't know why you guys are still using PP to compare 32M runs when the formula is skewed towards lower clocks hence if you run a 32M at 4Ghz with settings you are doing your benchmarks at 5.3 you will see a big difference in PP. You are not more efficient at 4Ghz, the formula doesn't work right.

[Gautam]

No dino, YOU don't work right.

[dinos22]

hahah i was waiting for you to chime in and give me a hard time about the formula but you know what i mean

i better rephrase it before Gatman police starts swarming all over this formula argument :p, the formula works but it has no practical use among overclockers who are unaware of bias when comparing runs on different clocks, particularly when the settings are not matching. Comparing fixed frequency CPU clocks works well as does keeping identical settings while ramping bclock which in term ramps up all the variables which are improving performance and maintaining the efficiency at high clock but a joe bloe overclocker using PP to compare his amazing 4Ghz run to a world record done at 7Ghz and saying he has better efficiency makes it impractical

ok go away now

[massman]

We all know the only way to compare efficiency is by beating someone's score at lower CPU frequency. Also, I'm not at 4GHz.

Here's a list of Clarkdale efficiency runs in HWBOT:

22348 - uncle fester ( Core i5 655k @ 3996.00 )
22520 - splmann ( Core i5 650 @ 4000.00 )
22710 - sup3rfly ( Core i5 661 @ 3996.40 )
23010 - trans am ( Core i5 655k @ 3990.00 )
23045 - TerraRaptor [benchmark.kz] ( Core i3 350M @ 2261.40 )
23059 - injenegr ( Core i3 350M @ 2261.50 )
23132 - SniperOZ ( Core i5 670 @ 3990.40 )
23242 - Redmax ( Core i5 661 @ 3996.00 )
23258 - pro ( Core i5 655k @ 5388.00 )
23293 - Massman ( Core i5 655k @ 5221.80 )
23325 - PlaSMaN ( Core i3 350M @ 2266.00 )
23349 - Massman ( Core i5 655k @ 5276.00 )
23383 - chew* ( Core i5 655k @ 4816.00 )
23397 - kikicoco1334 ( Core i3 350M @ 2133.80 )
23406 - Massman ( Core i5 655k @ 5217.00 )
23441 - TinTin@OCQQ ( Core i5 661 @ 5009.80 )
23545 - Ross ( Core i5 670 @ 5315.00 )
23561 - Badmuts ( Core i3 350M @ 2260.00 )
23572 - Flavius ( Core i3 530 @ 3996.00 )
23587 - ZoLKoRn ( Core i5 655k @ 5040.00 )
23595 - Ross ( Core i5 655k @ 5628.40 )
23610 - SchnickNick ( Core i3 350M @ 2261.00 )
23625 - bmwbaxter ( Core i5 650 @ 3200.00 )
23628 - -tze- ( Core i3 530 @ 4013.30 )
23630 - Jagvar ( Core i3 350M @ 2128.00 )
23640 - Bullet ( Core i5 660 @ 5017.00 )

[massman]

I'm sure tomorrow-night I'll produce a more competitive run in terms of end result. Efficiency-wise, unless I find the magic key to unlock DDR3-2100, it will be a no-go.

DDR3-2100! Too bad I had to ruin it with a copy-waza. Won't nail the M3F run, but low 23 is possible.

[Blue Storm]

nice runs PJ...I wish I had in my 655K memory controller as good as your 655K

[[XC] Oj101]

I don't know why you guys are still using PP to compare 32M runs when the formula is skewed towards lower clocks hence if you run a 32M at 4Ghz with settings you are doing your benchmarks at 5.3 you will see a big difference in PP. You are not more efficient at 4Ghz, the formula doesn't work right.

Because the clocks are so close, there's no 1.4GHz difference here, not even 5% in fact

[dinos22]

5% is a massive WTF are you on about

[[XC] Oj101]

I said NOT EVEN 5%. It's not even 2%.