Nehalem,One cpu to rule them all!

[cryomonkey]

I've been following your replies to threads in the news forum waiting for your report and you don't disappoint very nice

[Charles Wirth]

Good read Dave,

[The Warlock]

Good lord,Movieman, nice result, i understand why have fallen in love with it.
I have ordered a core i7 920 myself,the cheapest, but gee, what prices on those boards, hope the ram will be cheaper.

[naokaji]

The Emperor has landed.


I go over I7 and will it to produce some numbers...

(in a few weeks when it actually arrives in the uk).

Nice numbers Dave

[123bob]

Good read Dave,

..And your threads were not slackers either, Sir.... I made the mistake of finally downloading our own XS super pi, in order to see where I stacked up....Big mistake, now I have to call my therapist, Dr. SiG and report a severe case of pi inadequacy.. My best time, on my fastest, and only, harper rig @3.6g (crunching stable), at 1M digits was 12.844. A far cry from MM's results and very far away from yours.

Did you happen to do any power draw type analysis on your rigs? Particularly at clocks that us mere mortal crunchers (Read as, air or water cooled.) can run 24/7 on? I am particularly interested in this since I have to amortize/justify the total investment over a couple of years when I'm planning for the farm. If these are more cost effective when I pay the power bill, I can spend a bit more upfront, get more crunching done, and contribute my cash more smartly to our mission here.

We are also interested in your impressions on how "baked" you think the mobos you ran on are. Dave's board seemed pretty solid. You seem to imply a preference for your gigabyte board, at least I saw more screenies from it.

From what I'm seeing from all of you testing is that we are ready for the release of this beast on the computing world. Yes?

Thanks much, hope to meet you in Vegas this Jan.

Regards,
Bob

[emoners]

Thanks Dave for the info!

[naokaji]

Madshrimps has some power consumption numbers in their review, I7 uses only roughly the same as a E8600
linky

[emoners]

Madshrimps has some power consumption numbers in their review, I7 uses only roughly the same as a E8600
linky

LegionHardware on the other hand shows these i7s are way better at idle

[BlindFreddie]

Great stuff, MM. I've been wondering & waiting for this info ever since reading your post on WCG re. disguising the real ID of a CPU. You said then it was "a beast", but I was skeptical because Intel were only claiming about 10-15% over Penryn, clock-for-clock, and I thought that may have been mainly on apps with high data thruput, such as servers, where the 3 mem channels would kill the old FSB. I hadn't figured on HT being so effective.

If you really want to see the WCG points roll in, you might try crunching some FAAH WUs, or DDDT when it recovers from Hurricane Ike. With "single redundancy", points awarded are based on the benchmark internal to each WU, and they are about 13% higher than claimed by my Intels, and 10% below the greedy claims of my A64.
- BF, aka rickjb on WCG -

[Edit]: Re WCG points and FAAH, I just remembered that I'm running Xp-32 vs your XP-64. 64-bit BOINC inflates its points claims, because the 64-bit BOINC integer benchmarks are much higher than the 32-bit values. Your FAAH points may be scaled down, not up.
You might also be interested in my post #63, below.

[D_A]

... I7 uses only roughly the same as a E8600

... and is HOW many times faster?

[BlindFreddie]

MovieMan & All: I've been doing some thinking about those Nelly WCG benchmarks and have figured out some speed comparisons against my Yorkfields.

Caveats:
Both of my Yorkies are running XP-32 Pro SP3, so they are at a disadvantage to MM's XP-64. I believe that there is some difference in CPU-time per WorkUnit between XP-32 and XP-64, but I have not seen credible numbers on this ***See edit below. (For those not familiar with WCG, all the science applications are 32-bit, so run-times should be similar for both OSes, even if the Grid Computing Manager (BOINC) is 64-bit). There is a known large difference in benchmark scores between 32-bit and 64-bit BOINC clients in the integer (Dhrystone) scores, though not the floating-point (Whetstone) scores****See edit below. These scores are used by the BOINC client to determine the Points Claimed from WCG, and they affect the Points Awarded by WCG by different degrees on different WCG projects. The project most affected is HCC, the one that MM has crunched. His WCG Points Awarded will be inflated compared to mine, so we can't use them in real crunching-speed comparisons. However, it's interesting to have a look at the BOINC benchmarks.
My machines are:

• Q9650 @ 467 x 9 = 4203MHz = 1.114 x 3773Mhz (i7 speed). BOINC scores (fp/int) 4442/9101. Scaling this back to 3773Mhz gives 3945/8083. I don't have any HCC times for it as it's currently running 100% FAAH. *****See edit below.
• QX9650 @ 392 x 10 = 3920MHz = 1.050 x 3733MHz. BOINC scores 4146/8683. Scaling to 3773MHz gives 3948/7983. This machine is currently running HCC.

The scaled BOINC scores for my 2 machines are very similar, but they are not comapable to those for the i7 because of the 32-bit/64-bit issue.

Now let's examine the HCC execution times from my QX, which will have a disadvantage of 1% due to XP-32. Average CPU time for the last 6 WUs was 2.66 hrs. Scaling up by the 1.05 clock speed ratio gives 2.79 hrs. The i7's times running without HT are close to 2.5 hrs, giving a speed ratio 11.6% faster than the Yorkie (10.6% after 32/64bit). This is right in the range of Intel's claim of 10-15% clock-to-clock advantage to Nehalem.

Finally, the Hyperthreaded scores. MM's i7 does HCC WUs in about 4h24m, ie 4.4 hrs, times 8 virtual cores, equivalent to 4 cores in 2.2 hrs. The speed ratio is now a 27% advantage to Nehalem (26% after 32/64bit).

Yorkie is done but not disgraced by Nelly.

[Edit]:
*** I did some testing of XP-32 vs XP-64: I made a snapshot copy of my BOINC data directory, under XP-32, while it contained a mix of WUs from FAAH, HCC, HPF2. By selectively suspending processing of some WUs, I forced 1 WU of each type to start simultaneously, and ran, with "network activity suspended", until the longest of the 3 finished, and recorded the CPU times. I swapped my boot drive for one with XP-64 on it, copied in the BOINC data snapshot, and repeated the run. CPU times showed about 1% gain for XP-64 over XP-32 for all 3 projects.
[Before making my snapshot, I suspended crunching activity, suspended network activity, did Advanced >> Shutdown connected client, then exited BOINC manager. That removes any lock files.]
**** I found that the FP scores were much lower on XP-64, while the in scores were higher. Sorry, but I have misplaced the actual values.
*****I ran some HCC on the Q9650 @ 467 x 9. CPU Times scale with clock speed to the QX.

[WesM63]

Nice review MM!

I need 2 of these, now!

[Drwho?]

I am Francois, and I approve this message.
(Private joke for American's )

Core i7 on seti here : http://setiathome.berkeley.edu/top_h...edit&offset=20 the Who? machine! going up, in top 40 right now, I hope to get to top 5 with a single processor.

[Jaco]

I'd like to see the impact of memory timings with the on-die memory controller.

Tony hinted the best performance is with low latency timings like 6-5-5 or 6-6-6

[Martijn]

That is simply amazing... Dave, will we see 7M+ in the stats tomorrow?

[Bobsama]

Wow Dave. Awesome rig--but do get out your KillAWatt! I want to see load and idle draw.

[v0dka]

The first 8 WU I ran in 2 groups of 4 with HT off and as you see they finished in app 2 hours 28 minutes.
The last group of 8 WU I ran at one time with HT enabled and was happily surprised to see that they finished in less than twice the time of running 4 units with HT off.
4 hours 25 mins app .This shows the effect of HT and that it is a positive one.
From the brief numbers I can see it looks like this machine at 3733mhz will pull just over 20,800 points a day with HT off and well over 30,000 a day with HT enabled.

Wait..

2:28 (148 minutes) for 4 WU's
4:25 (265 minutes) for 8 WU's

Is that what you are saying?

So basically 37 minutes per WU without HT and 33.125 minutes per WU with HT. That's a difference of 11%

How is the output 20800 versus 30000 then? That's almost 30% more?

[NaeKuh]

Wait i just got the title now

dave the title one nehalem to rule them all is a joke at hera huh?

Wait til hera gets setup!!! GRRRRRR!~!

[CyberDruid]

Thanks Dave.


Go get him Naekuh.

[Movieman]

Wait..

2:28 (148 minutes) for 4 WU's
4:25 (265 minutes) for 8 WU's

Is that what you are saying?

So basically 37 minutes per WU without HT and 33.125 minutes per WU with HT. That's a difference of 11%

How is the output 20800 versus 30000 then? That's almost 30% more?

No, You misread. With HT off it does 4 WU at a time and completes the 4 in 2 hours 28 minutes.
With HT turned on it does 8 WU at a time and completes the 8 in 4 hours 25 mins.
Output was based on claims made and credit given.

Wait i just got the title now

dave the title one nehalem to rule them all is a joke at hera huh?

Wait til hera gets setup!!! GRRRRRR!~!

No, not at all my friend. Was based on the Nehalem being the best cpu in terms of computational power that you can buy.

Thanks Dave.


Go get him Naekuh.

I'm sure he will.. He has more $$ than I do!

To Blindfreddie:
Many thanks for the data.
I agree, Yorkies aren't dead.
With all the data what app I'm really glad I ran was Folding at home.
No benchmarks that could interfere with the numbers just turn it loose and see how fast it completes.
When I saw the machine capable of making almost 5000 points a day at 3733mhz vs RoadRunners 4GHz Yorkie making 3300 that told me what kind of power was in this thing.
BOINC and it's benchmarks have always been flawed but if you ignore those benchmarks in BOINC and just look at times to complete you see the true power.
8 HCC WU completeing in 4.5 hours is inline with my 8 core clover at 3000mhz but done at 3733mhz on just 4 cores. Perhaps that is a better way to understand the power that resides in these Nehalem cpu's. That's not scientific but more of a real world way of seeing into what it's truly capable of.

[v0dka]

No, You misread. With HT off it does 4 WU at a time and completes the 4 in 2 hours 28 minutes.
With HT turned on it does 8 WU at a time and completes the 8 in 4 hours 25 mins.
Output was based on claims made and credit given.

That's what I said. 4 WU in 2:28 and 8 in 4:25.

So 148 minutes for 4 WU's (or 296 for 8 it's the same) and 265 minutes for 8 WU's. Difference is 11%.

Based on the fact that there was 11% more WU's crunched or 11% less time needed to complete the same number of WU's I'd say that the difference in credit granted can never be larger than 11%. What you witnessed was probably quorum related.

Not that this isn't great, I mean, for a CPU that has only 4 real cores any bonus from using these four emulated cores is pretty much mindblowing. And if someone could explain that would be great. But 30.000 (or two Q6600) is just way too much.

[D_A]

Dave, could you grab us a screenshot from your WCG page showing the stats for this machine? I don't think I saw one before.

[D_A]

vOdka, based on the numbers we've seen so far that thing (with HT running) should be claiming over 43k (WCG that is) per day. It won't get it due to the quorum, but that's the expected claim.

[Movieman]

I am Francois, and I approve this message.
(Private joke for American's )

Core i7 on seti here : http://setiathome.berkeley.edu/top_h...edit&offset=20 the Who? machine! going up, in top 40 right now, I hope to get to top 5 with a single processor.

Francois: We really need to get you onto WCG. SETI is Ok I guess but if ET wants to find us I'm sure he will.
Finding cures for Cancer and Aids to me is all based on how much computational power we can throw at the problem.
With your personal and corporate resources we could cut the timeframe.
Think about it please.

I'd like to see the impact of memory timings with the on-die memory controller.

Tony hinted the best performance is with low latency timings like 6-5-5 or 6-6-6

Memory has always been my weakness so perhaps with some sharp guy like Tony I could get better performance out of this.
I wonder what his telephone number is?

Dave, could you grab us a screenshot from your WCG page showing the stats for this machine? I don't think I saw one before.

They don't mean anything yet, wait a week and then we'll have good info.

vOdka, based on the numbers we've seen so far that thing (with HT running) should be claiming over 43k (WCG that is) per day. It won't get it due to the quorum, but that's the expected claim.

Pretty close, I saw app 40,800 as the claim but remember, BOINC NEEDS 5 solid days history with a machine to see that the claims are consistant OR it automatically takes the other machine in the quorum as the norm and you as the outliar.
Now for today, the first "full" day just running WCG I see this for the 1/2 day update..

Statistics
Date.... Total Run Time Points Generated Results Returned
11/3/08 0:005:00:06:49 16,356...................... 28

based on the above being credit granted for essentially 5/8 of a full 24 hour day those numbers extrapolate to a granted credit of 26,170 points per day but remember, the machine is being killed in the quorum so I only expect those numbers to go up.
Even if they don't it is by far the best performing quad that we've ever seen.

[D_A]

There is another little thing about my projected claim. It assumes 24hrs of work. Every time the system completes a unit and grabs another it loses a couple of seconds, so unless the units take longer than 24hrs each to process there's some lost time. More units processed means more lost time but faster subsystems (using SATA2 instead of old IDE HDDs for example, but also memory speed etc) reduce that amount of lost time.
I'm only talking seconds here, maybe one or two between units, but it adds up. Mind you, even if you doubled the total speed of all your subsystems you'd only gain a matter of a few seconds, maybe a minute, worth of actual work time ie stuff all over a day. I think (personal opinion) this is why there's no perceivable difference between single and multi-channel RAM *for this specific purpose*. Even if you could cut that delay down to a couple of milliseconds, you'd only gain a couple of minutes over a whole day.