Crunching comparison 2600K and 3770K

[OldChap]

Update the software and ran them side by side ....No difference in temps shown on either machine. Maybe GT is just wider for more cores?

[OldChap]

Some Faah numbers added to first post.

Variable type work so not sure if we can take anything from these.

[WFO]

When you get done with the current battery of tests @ 4700, perhaps you can drop the v-core to stabilize clocks to 4600? I currently have 1 3770K @4500, 1 @4600 and 1 @4700. The one that got bumped to 4700 is stable but has been producing about 300 ppd less than the lower clocked chips. There appears to be no difference in points with the same projects between 4500 and 4600. Oddly the chip at 4500 needs the most v-core for stability. 1.240 It also has the worst cooling. ( Coolermaster Hyper 212 EVO) Still gathering data.

In case anyone is wondering, 4600 is taking 1.232 with the Thermalright Silver Arrow. 4700 is taking 1.224 with an Apogee HD and Alphacool NexXxos UT 60 240 full copper rad.

Edit: Of course as soon as I posted this, the highest clocked rig put up 54K to 48K to 47K. MY IBs all have DDR 3 1600. Soon to be remedied. It's looking like I'll have to wait 30 days and compare RAC.

[OldChap]

A lot of comment I have seen leads me to believe that to some fairly important degree temperature plays a part in what is in fact a stable voltage.

For mine;

1.248v = stable LinX at around the 110 GFlops level
1.26v = stable Prime95 6 hours
1.27v = 1 restart crunching mixed wu's in 24 hours
1.28v = stable mixed and now the first 2 wu types for 4 days.

I need to let the rig get to a position in the chart at this speed and, as the two will be doing the same wu types for that period (they are on the same profile) I believe that too will be a good indicator. Once I am in that position I can drop the clocks as you requested if you wish but I feel you might get some mileage out of doing a similar test.

Put yours on the same profile, run a couple of days of each project then change to a new one. By the end of around 2-3 weeks you should be in a position to see what is working.

Incidentally I have my internet on for 3 hours a day now in an attempt to see if I can also see how many wu's complete in that 21 hour period for each machine. Probably of better use for Cancer wu's.

We can discuss your requirements some more at the end of this sequence of testing.

[OldChap]

50 results compared for HCC in post 3.

[Conquistador SW]

I have exactly the same experience as you. My chip is Linx stable at 1.25v, but only at 1.29v it is 100% stable. I found that a nice way too find out if you are completely stable is looking in windows "event viewer" and then under "custom views" -> "administrative events". If you see some WHEA-logger warnings there, it is probably not stable, even if Linx or other tests say so.

And I think you are right that temperatures influence the stability. I de-lidded my 3770k and it is stable at 4.7Ghz 1.29v 67C while before it was not 100% stable at 4.7Ghz 1.29v 88C.

[OldChap]

50 results compared for HPF2 in post 3.

Although bigger this seems the better format. Let me know if it needs adjusting, if having the data rather than just the results serves any purpose... stuff like that

[OldChap]

Question; Need to crunch WU's that use a lot of memory during this test.

In view of the fact that the 2600K rig is only fitted with 4Gig should I save those tests until the end, then install 8Gig the same as in the 3770K or is there another better way???

[haschioz]

that depends how much the wu require. you can run only 2 wus and find out how much they take if it is ram limited safe it for last and then run it with 8 gig. for approximation use the wu that uses most ram.

[OldChap]

Well, even though they tell us that clean energy requires a bunch of memory it seems that it is not the only one....

CEP2 Mem =1,024 MB Disk = 2,048 MB
DDDT2 Mem = 1,024 MB Disk = 250 MB

Maybe I should run 2 tests with the 2600K; 4G and 8G

[itznfb]

So I now have 10 days of solid uninterrupted data from my 2600k @ 4.0ghz 1.225vcore.
Average of 150 watts from the wall, average of 61c/26.7c ambient, with an average of 41690 points per day and average of 56 results per day

My 3770k is now in the same machine at the same settings under the same cooling. I'll let it run for about 17 days and compare the last 10 days.

[Otis11]

When you get done with the current battery of tests @ 4700, perhaps you can drop the v-core to stabilize clocks to 4600? I currently have 1 3770K @4500, 1 @4600 and 1 @4700. The one that got bumped to 4700 is stable but has been producing about 300 ppd less than the lower clocked chips. There appears to be no difference in points with the same projects between 4500 and 4600. Oddly the chip at 4500 needs the most v-core for stability. 1.240 It also has the worst cooling. ( Coolermaster Hyper 212 EVO) Still gathering data.

In case anyone is wondering, 4600 is taking 1.232 with the Thermalright Silver Arrow. 4700 is taking 1.224 with an Apogee HD and Alphacool NexXxos UT 60 240 full copper rad.

Edit: Of course as soon as I posted this, the highest clocked rig put up 54K to 48K to 47K. MY IBs all have DDR 3 1600. Soon to be remedied. It's looking like I'll have to wait 30 days and compare RAC.

:ehhm: Talked to a little birdy about that - said that chips (he thought since Nehalem, but not 100%) will actually put unused sections to sleep when a parameter gets too far outside of it's limits, and only wake back up when something needs it. Because of that, even though you can use the whole chip, you waste cycles while it starts up. Could be caused from volts being too low or temps too high (even if your cpu temps as a whole are good-the ones you can measure using software, certain internal temps might be too high-temps you don't have access to monitoring.)

Edit: and temperature has a significant role in stability at given voltages... I don't have my book on me (left it in Austin) but for those interested higher temperatures cause vibrations (called phonons), which in-turn cause more scattering of the electrons passing through the transistor (basically, more vibrations means more collisions slowing the electrons down as they travel). This causes the circuit to transition more slowly - and if it is too slow for the given clock, you get an error. Raising the voltages makes the electrons travel faster through the transistor by giving them more initial energy. Because of this, the circuit transitions more quickly. The problem arises when these high energy electrons do happen to scatter - they release this excess energy as heat, causing more phonons and increasing the effects of scattering in a downward spiral. At our points in volts/temperatures/material parameters, we are trying to overcompensate for the additional clock speed by giving the electrons significantly more starting energy than they are supposed to have in the designated transistor. This causes a substantial increase in scattering through increased temperature, which we then try to correct for with our massive coolers. If anyone's actually interested let me know, otherwise i'll end my rant...

[OldChap]

HFCC added.

[OldChap]

C4CW added

[OldChap]

Waiting for a bunch of validations just now so just for the record I got a few CFSW betas on each machine .... not enough for a page print but completion times were 0.62 and 0.63 for the 2600K and 0.59 for the 3770K so again falling around 5% advantage to the IB cpu

Wfo; did you do some tweaks? I see your higher clocked rig improving and the whetstone numbers up a bit

[itznfb]

So in my testing with everything being the same except the CPU installed the 2600k produces just over 1000 points per day more than the 3770k. 1000 ppd is most likely within a margin of error in testing. The 3770k did consume 20 watts less than the 2600k while producing about the same output. One thing to note is that it is more likely to find a 2600k or 2700k that will clock much higher than the 3770k at the same voltage. So it still may be advantageous to go with the 2600k/2700k being that it will be significantly easier to cool and most likely clocking much higher.

Another note.... I was running a 2500k @ 5.0GHz and it was not able to keep up with the 2600k or 3770k @ 4.0GHz. So a 1000MHz clock advantage didn't make up for the lack of HT. I thought that was interesting.

[haschioz]

itznfb, not keeping up in ppd or results?

[itznfb]

itznfb, not keeping up in ppd or results?

Both. Obviously the runtime on each individual job is completing faster but apparently not fast enough to make up for the additional 4 threads with the 2600k or 3770k. The HT chips are completing more jobs and resulting in ~4000-5000 more ppd. Pretty impressive for HT technology.

[OldChap]

The 3770K has had some issues with GFAM (invalids) and today two errors on CFSW ....

Got to the bottom of the problem (I hope)... Had set a voltage within windows then not transfered it to BIOS after testing.... Thank goodness for taking copious notes

The new type CFSW WU's sure do complete fast (under 40 mins) so if you are looking to increase your number of wu's completed total.....

I believe there are some new ones for other projects too (GFAM, SN2S, DSFL) so perhaps it is better that those results from the older type have not been charted yet.

[OldChap]

After a bunch of issues that it seems were not all my fault I have revamped post 3 to summarise what I have so far.

There have been projects with no work, projects changing their app version, in fact you name it I seem to have run into trouble with it. You only have to see what it did to my 2600K averages to realise this has not been smooth sailing

I need to let things settle for a while before I attempt the last few projects.

PLEASE consider this a guide only.... You will need to confirm on your own rigs that these results hold true for you too