i7 920 vs i7 860 direct test comparison

[Otis11]

The waiting option: XS style

Buy one now, and if you were wrong just buy the other one too. Don't worry, they'll both crunch.

[drizzt5]

So I am guessing the 920 is the better option....

[123bob]

So I am guessing the 920 is the better option....

It just depends on your situation. What this thread shows is that at least you won't get hurt in crunching performance, clock-for-clock, by going 860. The rest depends on what the machine is for, your immediate budget, your power cost, and the upgrade path considerations...

Bob

[123bob]

The waiting option: XS style

Buy one now, and if you were wrong just buy the other one too. Don't worry, they'll both crunch.

I'm thinking "XS style" is the way Emu and I did it. Get one of each right away so you can "evaluate" them for the team....

Bob

[Otis11]

I'm thinking "XS style" is the way Emu and I did it. Get one of each right away so you can "evaluate" them for the team....

Bob

On second thought, I like yours better.

Glad you pointed that out!

[123bob]

As I said above, I wanted to see how close identical parameter runs of our new test method would come to each other. So, I ran another test on the 920, using EXACTLY the setup from the previous run.

So, this post is about how accurate we can expect the data, using this method, to come in.

Here's the previous run posted about above.



As reported before, it's average is 3.9530787 Hours/WU.


Here is the raw data from the next run, at the same exact settings.....



It comes in at 84:659:780.

Converting to hours = 95.2000000 hrs. (Yup, I know it seems strange that it hit zeros in the second decimal place, did the math several times....You're welcome to check it....)

Dividing by 24 WUs gives 3.9666666 hours/WU


The difference between this and the run above it is .013587962 hours, or 48.91 seconds.

So this second run was about 49 seconds SLOWER than the first run. Remember, exact same parameters were used.

So, I would put the variance of this test method at around +/- a minute per WU, or so, at least. Many more runs would have to be done to get a statistically valid varience, but I don't have that time since it blows 12+ hours of crunching away to do each run.

What this shows me is that if we get within a couple of minutes per WU for a test run, we can just as easy say it's equal, due to the test variance.

As we get finer into parameters, more runs would need to be done to establish averages. I'm thinking like at least 10 runs per setting change.... I'm not going to be doing that, and I would not expect anyone else to do it either...

However, it is clear that this test method is WAY better than we were attempting before.

So, here's the order of accuracy as I see it.

Most accurate - This method of running the EXACT same set of WUs.

Less Accurate - The method of taking a random batch of many completed WUs and averaging completion times.

Least accurate - Taking daily scores over a very long time, say several weeks, and working those backwards.

I'm not bothered by this in any way since the nature of what we are doing is probabilistic. As I understand it, a given WU is calculated down a path and has many branches. It may not take the same branch every time it's run. Therefore, we should not expect it to complete in EXACTLY the same time.

Agree?

I've been dying to get into this conversation years ago. I just never had a real reason. Damn my engineering genes.....

Regards,
Bob

EDIT: Curious side note. It took exactly 13 hours, 5 minutes of "clock time" to complete this run. I don't think that means anything, but I was curious to see, so I timed this run. I'll bet this number will vary even more than the run-time data. :END EDIT

[emuexport]

First of all bob, let me thank you again for your time and dedication put into this, truly amazing

Below is just my 2 cents on your conclusions.

I agree, they are equal crunchers, as was expected. The 860 may be more power-efficient at 3,7Ghz, however, judging from the LGA1156 and 1366 CPUs I have tried until now, this picture changes dramatically if you go to around 4Ghz. Almost all Lynnfield CPUs seem to require vast amounts of Vcore to handle 4Ghz, 1,4V seems like average, while most D0 920's can do 4Ghz at around 1,25V. I compared an X3450 running 4Ghz (bad one.. 1,43V and crunching at 87C for this test ) to an average i7 and came up with ~44W less for the 920 with otherwise identical setups (PSU, GPU, SSD identical).

Hey mate I'm dedicated too just more distracted

Not sure wether they are equal crunchers the higher you go though but I am sure you will confirm my results below shortly. I greatly appreciate your knowledge in these as I am just bashing my way through LGA1366 and LGA 1156 atm

I'm also glad to find out that my 860 isnt such a dog after all. When shes done her job she'll go back to more appealing voltages if she wants to.

Yes, using a high QPI mult seems to be crucial for Lynnfield performance. However, I could not detect the same behaviour with Bloomfields, I ran my 975 at different BCLKs and QPI mults over the last half year or so, and I never noticed any changes in PPD or WU runtimes caused by different QPI bus speeds. Of course I didn't do a real scientific test, but I still don't think Bloomfield cares about high QPI mults the same way Lynnfield does. maybe someone wants to try it out?

hmmm

Yeah, most probably it is the higher uncore speed giving the boost, as 1T vs. 2T does next to nothing on Intel X58 chipsets, not even if you run Everest mem benches etc.

hmm more tests required at 4GHz me thinks.

The results are in the above post. It seems to me that QPI makes a pretty good difference. The Uncore/NB difference is there, but not as pronounced.

I figure if we know this, we all can tweak for it and get the most out of our rigs. I would suggest none of us go "ape crazy" to push these, just be aware that they should not be ignored.

Should check out my results for more confusion mate!

Your tweaks look good. I realized after I blew all the WUs out of the 920 (or "target" rig) that I could have just moved them out for a day and put them back. That's a good improvement to the original write up.

I disconnected the net cable just to be REAL sure my test WUs didn't escape on me... I had the horrible picture of them escaping from the 920 rig, getting out in the wild and possibly bringing the server down, or destroying the planet..... I've since had the chance to see how "Net Act Supend" works and will likely do that in the future.

Regards,
Bob

Yeah I found out about net act suspend when I was manually updating the farm via a shared 3G wireless card on my laptop. If I didnt suspend it the computer gradually ground to a halt as it was constantly trying to connect to the net.

That's an 8% clock difference, and since we know that performance per clock is the same also 8% greater production. For just a higher investment in hardware I'd say that is a nice bonus.

But then you have the power consumption. Personally I would always go for the 860 at a reasonable 3.7Ghz and save ~40w. Crunching is all about good performance for the lowest possible costs.

Also, it's not only the inital investment in the 860 system that is lower compared to the 920 system, I think you can increase that difference when you realise that a budget board and budget RAM are good enough because you are not going to push it past 3.7. The 920 pretty much needs high end stuff.

It will definetely be interesting once we put power costs against the 3.7GHz results that Bob has done.

I agree with you V0dka, for pure cruncher duty. In fact, I think INFRNL's choice may be a winner for that. I run my pure farm rigs so they don't need a babysitter. I want them to run, run, and just run. If I can have that and power efficiency, I'm interested.

However, for some folks using rigs for games, applications, and such, the 920, the 860, and budget come into play. I can see that aspect too.


In other news......Well, I tried my best to get 4 gig out of the 860. I couldn't even get 3.885gig stable on it. (185x21)

I tried the max vcore the bios would let me set. I mean I had CPUz showing 1.536v at idle! I could boot, but no primes....

I jacked up all the other voltages to what they had for max safe, dropped the QPI to x16, still no love. This particular chip ain't going there.....

I dropped it all back to the magic 3.7 point and it's fine. No harm done from the voltage, at least right now....

So, it's up to Emu to give us the 4 gig point. Best of luck to you my friend!

Regards,
Bob

Cant wait to see INFRLs results! (Might need another "upgrade" )

4GHz results incoming!

I'm thinking "XS style" is the way Emu and I did it. Get one of each right away so you can "evaluate" them for the team....

Bob

I thought it was dont tell the wife you bought another computer! Mine are all "upgrades" to my existing comps

[Aerou]

they put a frequency activated resistor into 1156 cpu, to destroy the idea of a perfect crunchers rig, so we still may drool at Jcool's 1366 under SS , j/k

[emuexport]

So here are the screenies.

i7 860 @ 4GHz


i7 920 @ 4GHz


Well different WUs to Bob so the duration will be different but you get the idea.

The 860 @ 4GHz has a runtime of 99:51:44 for 24 WUs or 4:09:39 per WU on avg
The 920 @ 4GHz has a runtime of 96:09:33 for 24 WUs or 4:00:24 per WU on avg.

This is a 9min difference per WU or 4% per WU

So not sure what I need to change except for the memory settings. Any ideas as I'll set these to run again if required.

[jcool]

Nice work emu.. is that on the highest QPI mult already for the 860? If not, try increasing it.

[emuexport]

So you dont think the memory timings explain that difference then?

[jcool]

I don't think so, no. That 860 has more than enough bandwidth as it stands. Still, the observation that WCG doesn't care about ram performance is based on running solely HCC, where this holds 100% true.
Maybe HFCC is more memory intensive. What indicates HFCC being more memory intensive is the fact that on my Quad Opteron, HFCC is incredibly slow with the TLB fix turned on (24h WUs). With the TLB fix OFF, I get normal runtimes (8 hours), while the difference in HCC was only 30minutes or so.

What the TLB fix does is cripple the memory bandwidth of the CPUs by a huge margin.

So, maybe HFCC is memory-intensive.

[emuexport]

hmmm will have to see. I'll run these QPI queries you requested first then we can check out timings.

I have dropped the multi on both CPUs but upped the BCLK to 201 (mainly cos this EVGA board doesnt appear to have a QPI multiplier!?!?!?!)

Screens below:

[Vinas]

It may have already been said but increasing your PLL and IMC voltages seem to have an effect on performance. Give it a go maybe?

Also, thanks for the tests. I'd be interested to see the 920 bumped up in vcore to match the 860 (even if it's not required). This way we would have an absolute apples to apples comparison... Well as close as possible anyway. My suspect is that the 860 would still be a little faster and more power efficient. It is a newer design, after all so we can pretty much expect that.

Great work here, folks!

[123bob]

Thanks Emu!

You own the 4 gig point since I can't seem to get there with my 860...

It will be interesting to see how the QPI alone affects the runtimes. Now you have two rigs with just a QPI boost running. I'm expecting it to make a noticeable difference.

After this are you planning on pushing out the RAM timings on the 920 to match the 860? Or the other way around, reduce the RAM timings on the 860 to match the 920? I would think it would be easier to push out the 920 and still have a stable clock than to pull in the 860.

Hehe, maybe we should hit the team up for a couple of XS lab coats with all this science going on here....

Rock on Emu

Bob

[emuexport]

It may have already been said but increasing your PLL and IMC voltages seem to have an effect on performance. Give it a go maybe?

Also, thanks for the tests. I'd be interested to see the 920 bumped up in vcore to match the 860 (even if it's not required). This way we would have an absolute apples to apples comparison... Well as close as possible anyway. My suspect is that the 860 would still be a little faster and more power efficient. It is a newer design, after all so we can pretty much expect that.

Great work here, folks!

Will look into the PLC and IMC voltages at the next opportunity.

I understand where you are coming from but if my 920 is average in terms of voltage required to meet 4GHz then this is an advantage over the volt hungry 860. Besides I dont think I want to put 1.4 VCore on my 920 for an extended period

Thanks Emu!

You own the 4 gig point since I can't seem to get there with my 860...

It will be interesting to see how the QPI alone affects the runtimes. Now you have two rigs with just a QPI boost running. I'm expecting it to make a noticeable difference.

After this are you planning on pushing out the RAM timings on the 920 to match the 860? Or the other way around, reduce the RAM timings on the 860 to match the 920? I would think it would be easier to push out the 920 and still have a stable clock than to pull in the 860.

Hehe, maybe we should hit the team up for a couple of XS lab coats with all this science going on here....

Rock on Emu

Bob

Yeah mate I thought mine was a dog when jcool suggested less VCore. Have you tried MOAR VCore

Well the 920 crashed overnight with the increased QPI so I guess just adjusting that and nothing else wasnt enough Off to do my HUET today so no time this morning, it will have to wait till this arvo.

Memory timings is next on the list after this one. Then PLL and IMC as Vinas suggested then I'm going for max clock with stock VCore

Oh and I am only doing this for the t-shirts There is t-shirts right?

[DAK1640]

XS Lab coats in transit, complete with pocket protectors...

[123bob]

XS Lab coats in transit, complete with pocket protectors...

Have you tried MOAR VCore

Oh and I am only doing this for the t-shirts There is t-shirts right?

I gave it all the vcore my bios and board would allow. I had CPUz 1.536 vcore on it at idle for 3.885 gig and she still wouldn't prime.....It just won't go.

Dak has us covered. The pocket protector is a nice touch. Thx Dak! I haven't worn one of those in years....

Bob

[drizzt5]

I gave it all the vcore my bios and board would allow. I had CPUz 1.536 vcore on it at idle for 3.885 gig and she still wouldn't prime.....It just won't go.

Bob

Is that the situation with most 860's? Or do most do 4ghz? And 4% difference is not so bad.

[123bob]

IDK, we have exactly a pool of two to work with here, for now....Buy one, Let us know....

Bob

[jcool]

The X3450 ES I have right now can hardly do 4Ghz... not on air anyway.

[emuexport]

XS Lab coats in transit, complete with pocket protectors...

Sweet! I definetly need a pocket protector!

I gave it all the vcore my bios and board would allow. I had CPUz 1.536 vcore on it at idle for 3.885 gig and she still wouldn't prime.....It just won't go.

Dak has us covered. The pocket protector is a nice touch. Thx Dak! I haven't worn one of those in years....

Bob

So you arent hard modding the board with some solder? Why not?

I thought mine was a dud when jcool said they shouldnt need so much VCore but I guess you never know!

Is that the situation with most 860's? Or do most do 4ghz? And 4% difference is not so bad.

Mine is one water with a triple rad and fans at full (Scythe Gs) and a Fusion v1 with whatever TIM I could find ie not very good temps are around 60-65 DegC

IDK, we have exactly a pool of two to work with here, for now....Buy one, Let us know....

Bob

Yes Bob is right go down the store and get one mate!

The X3450 ES I have right now can hardly do 4Ghz... not on air anyway.

hmmm bummer anyone want to buy my "Golden 860 Chip" then?

In other news the 860 with increased QPI has finished its results but the 920 didnt make it through the night!
I'm currently running prime95 in Blend (as per jcools settings) and will run this batch of WUs ehn it passes enough.

My EVGA board only has the option of 4.2GHz and 4.8GHz QPI yet the 920 goes much higher. So this will be another thing to test at 4GHz and see what difference it makes for the 920.

Another update shortly. Spent half the day in class and the other crawling around smoke filled rooms and fighting fires. So I am off for a "cold one" (read beer!)

[DAK1640]

You two guys MADE me purchase an i860 with a Gigabyte mobo. Arrives today, so come Saturday we will have a pool of 3.

[emuexport]

well Bob "made" me buy one too!

[Aerou]

Careful, don't fry the 860, or your mobo socket .... http://www.xtremesystems.org/forums/...d.php?t=234723