Anandtech benches Nehalem

[Donnie27]

I think I am going to wait till after the big release to purchase, one, because I want to wait and see what a little maturation will do, as well as (hopefully) some price rudection, and seecondly I wont have the cash till then regardless. The real question is when to upgrade my graphics, before or after Nehalem.


The one thing that can be said about this chip, is it is a culmanation of all current cpu technology , regardless of who or where it came from, Intel got it on the die first.

It's already about 9 months old.

I disagree with most folks. Intel will launch a few None Server Nehalems, maybe the Top two models for the desktop.

2.66GHz one and 2.83 is shown on their Roadmaps for Q4-08.

[gallag]

QX9770 @ stock 3.2Ghz
8GB RAM @ 1066Mhz
Foxconn Blackops

its in my sig. it is NOT at all tuned atm due to problems with the MB.

anyway i did the math so that i could see what clock for clock difference would be.

the only two tests I duplicated were cinebench and pov-ray. all anand's other tests were proprietary and he does not give the info needed to duplicate them, but I saw the exact same result on pov-ray. anands penryn results were way too low on this one as well.

is pov-ray also 10% better on vista 64??

Man, if two different systems only have a cpu in common then there mobo, memory & hdd will be different and thats only the hardware, You then have software rangeing from difrent bios,drivers,operating system, apps, app revisions etc.

How you are stunned that any of this could cause a 10-20% discrepancy is stunning me.

[GoThr3k]

I tend to think the performance gap will be app dependant. These might range from 8% to 50%.

for single threaded?
50% is a lot you know

[villa1n]

"Cinebench shows us only a 2% increase in core-to-core performance from Penryn to Nehalem at the same clock speed. For applications that don't go out to main memory much and can stay confined to a single core, Nehalem behaves very much like Penryn. Remember that outside of the memory architecture and HT tweaks to the core, Nehalem's list of improvements are very specific (e.g. faster unaligned cache accesses)."

so basically its a penryn when its single threaded, but owns it up when multithreading and smt come into play.
From what i see, the logical cores then, along with the cache latency make the magic, but performance isnt really accross the board, but rather application dependant.
Still for server apps, and crunching / workstation type things, this is gonna be a BEAST. but for desktop, doesnt make sense atm, price wise.

[shiznit93]

where did you get the 2% number? the single threaded cinebench test shows 25% more performance at the same clock speed?

[zanzabar]

right now this is the $1200 cpu, so with the resonable priced consumer grade the core2 will still be faster all this shows is that HT works with encoding clock for clock, but i want to see how high it clocks if its not over 3.2-3.3 with air then it wont beat the core2 with its 3.8-4.0, and games dont use a quad now so using 8 cores wont come for at least a year or 2 more

where did you get the 2% number? the single threaded cinebench test shows 25% more performance at the same clock speed?

http://anandtech.com/cpuchipsets/int...spx?i=3326&p=7

thats 3015 to 2931 so thats 2.6%, and thats not impressive since it has 30-50% more memory bandwidth and the core2 is running with stock 333mhz fsb

[ceevee]

How you are stunned that any of this could cause a 10-20% discrepancy is stunning me.

I never said that. You are getting me confused with another member.

I have simply said that Anand's Penryn results were extremely low and not realistic.

and I'm not just guessing at this. I ran the same benchmarks and have seen first hand that his penryn results that he compares Nehalem to are not real world Penryn results.

-edit-

Nehalem will be a great upgrade and I will be one of the first to adopt it, however I take issue with the degree of improvement he is trying to show. It is going to be more like 10%-15%.

[savantu]

..

thats 3015 to 2931 so thats 2.6%, and thats not impressive since it has 30-50% more memory bandwidth and the core2 is running with stock 333mhz fsb

That's actually incredibly good.Why ? Because , IMO , Penryn has a better cache subsystem :

Penryn - 32KB L1 3 cycles , 6MB L2 15 cycles
Nehalem - 32KB L1 4 cycles , 256KB L2 11 cycles , 8MB L3 39 cycles

Nehalem has 4 cycles L1 , that's a lot , but if probably hidden with SMT and other techniques.The small L2 has very good latency , but at only 256KB it's really tiny.The L3 is very large , but also very slow.

Basically , it's far from optimal for single threaded apps ( Core/Penryn are best there ).

To be honest , I expected Nehalem to be slower than Penryn in single threaded apps that aren't BW dependant.This might still be the case , but it looks like Intel did its job.
Hell , even maintaining Penryn single thread performance coupled with K8/K10 scalability ( in fact even better ) makes Nehalem an excellent all around monster.

[villa1n]

where did you get the 2% number? the single threaded cinebench test shows 25% more performance at the same clock speed?

Right here.

[Metroid]

right now this is the $1200 cpu, so with the resonable priced consumer grade the core2 will still be faster all this shows is that HT works with encoding clock for clock, but i want to see how high it clocks if its not over 3.2-3.3 with air then it wont beat the core2 with its 3.8-4.0, and games dont use a quad now so using 8 cores wont come for at least a year or 2 more

http://anandtech.com/cpuchipsets/int...spx?i=3326&p=7

thats 3015 to 2931 so thats 2.6%, and thats not impressive since it has 30-50% more memory bandwidth and the core2 is running with stock 333mhz fsb

The article has changed since. This is not a $1200 CPU, certainly extreme versions will come at the beginning with a 3.3Ghz clock speed.

Interestingly enough, none of our standard CPU benchmarks are single threaded at all - even the most benign ones are multithreaded (including the games). I did run some single thread Cinebench numbers though:

Nehalem - 3015
Q9450 - 2396

Compare the blue and yellow graph to Anand's two graphs. According to these benchmarks, "old" Penryn beats "new" Penryn by about 38% in single-threaded Cinebench and 17% in multi-threaded Cinebench.

http://images.anandtech.com/graphs/n...0043/17014.png
http://images.anandtech.com/graphs/n...0043/17023.png
http://images.anandtech.com/graphs/a...1132/16054.png
http://www.anandtech.com/printarticle.aspx?i=3153

A mature Penryn system should score closer to the 3000 mark then what Anand listed.

You can look at other review sites as well
http://www.hardwarezone.com.my/artic...521&cid=2&pg=8
http://www.overclockersclub.com/revi...el_q9450/7.htm

This should be raising some red flags people

Fixed.

That was entirely an error on my part, it wasn't a SP1 or a configuration issue. It was an Excel spreadsheet malfunction I used data from the wrong column (first run data vs. average run data) for Cinebench. Everything else looks to be exactly where it should be but I'll make another run through the spreadsheet to make sure.

I just reran the numbers to confirm and now things make much more sense. Not only are our XCPU scores virtually identical to what they were for the Phenom article, but the single threaded tests make a lot more sense. Furthermore, the scaling from 1 to n-threads makes a lot more sense now too. Penryn gets a 3.56x speedup from multithreading while Nehalem gets a 4.18x speedup - the difference in scaling partially being due to HT.

Thanks for bringing this to my attention and sorry for the mixup.

Take care,
Anand

Some tests are controversial at some point as we can see by how many times they have been edited but the truth will come soon, Anantech is a great website reviews that tells most of the time the truth despite minor mistakes that happens here and there.

Metroid.

[GoThr3k]

That's actually incredibly good.Why ? Because , IMO , Penryn has a better cache subsystem :

Penryn - 32KB L1 3 cycles , 6MB L2 15 cycles
Nehalem - 32KB L1 4 cycles , 256KB L2 11 cycles , 8MB L3 39 cycles

Nehalem has 4 cycles L1 , that's a lot , but if probably hidden with SMT and other techniques.The small L2 has very good latency , but at only 256KB it's really tiny.The L3 is very large , but also very slow.

Basically , it's far from optimal for single threaded apps ( Core/Penryn are best there ).

its actually quite bad... you forget the memory access latency in your comparison
not to forget all of those cycles are different depending on how large the blocksize and stryde is

[savantu]

I never said that. You are getting me confused with another member.

I have simply said that Anand's Penryn results were extremely low and not realistic.

Allow me to doubt this.I've looked at other reviews and they all the the Q9450 at around 10200/10500 points in Cinebench 10 32bit.

http://www.hardwarezone.com/articles...521&cid=2&pg=8

[zanzabar]

That's actually incredibly good.Why ? Because , IMO , Penryn has a better cache subsystem :

Penryn - 32KB L1 3 cycles , 6MB L2 15 cycles
Nehalem - 32KB L1 4 cycles , 256KB L2 11 cycles , 8MB L3 39 cycles

Nehalem has 4 cycles L1 , that's a lot , but if probably hidden with SMT and other techniques.The small L2 has very good latency , but at only 256KB it's really tiny.The L3 is very large , but also very slow.

Basically , it's far from optimal for single threaded apps ( Core/Penryn are best there ).

To be honest , I expected Nehalem to be slower than Penryn in single threaded apps that aren't BW dependant.This might still be the case , but it looks like Intel did its job.
Hell , even maintaining Penryn single thread performance coupled with K8/K10 scalability ( in fact even better ) makes Nehalem an excellent all around monster.

i want to see duel and quad benches and not ones that will do 8, or see HT disabled, this seams like smoke and mirrors to me. this just reminds me of the socket A to 754 were there was no advantage to the 754 but people clamored to it then it was fixed with the 939, and i still dont trust HT


as for the benching tri channel and better core to MC bandwidth should balance the cash, but dosnt the c2 have better cash in it and thats out or about to be out, and when u oc the core2 it scales but the NH dosnt look like it (and they didnt change the fsb for the 2.93 so thats questionable to making this marketing and not benching)

[Loser777]

Damn, why do I always choose to the worst times to build PCs?
Right when I finished my Dual-Core Athlon rig Core 2 came out.

I was dead set on getting a Q9450, and this has to rear its ugly ass.
I don't think I can wait to Q4. Plus, I'm willing to bet that a Q9450 @ 3.8 can beat this at stock, which is good enough for me...
I wonder if the IMC will be as crippling as it was to A64s OCing.

I'm thinking that a Nehalem + Mobo will cost around $1500 - $1700.

[perkam]

Damn, why do I always choose to the worst times to build PCs?
Right when I finished my Dual-Core Athlon rig Core 2 came out.

I was dead set on getting a Q9450, and this has to rear its ugly ass.
I don't think I can wait to Q4. Plus, I'm willing to bet that a Q9450 @ 3.8 can beat this at stock, which is good enough for me...
I wonder if the IMC will be as crippling as it was to A64s OCing.

I'm thinking that a Nehalem + Mobo will cost around $1500 - $1700.

Unless ofcourse Intel will make these 2P and 4P exclusively at launch, you should be able to get a Nehalie for $999 and a mobo for $250 so $1,250 is a much better estimate.

Perkam

[Loser777]

I'm not willing to spend more than $350 on a CPU at MOST...
If a Q9XXX can beat a stock Nehalem when OC'd then I'm pretty satisfied because Nehalem won't have the OC headroom.

[savantu]

I'm not willing to spend more than $350 on a CPU at MOST...
If a Q9XXX can beat a stock Nehalem when OC'd then I'm pretty satisfied because Nehalem won't have the OC headroom.

You know that how ?

Ah yes, from Fuad.Don't forget the mountain of salt , though...

[Metroid]

Intel will not launch only extreme versions of Nehalem on Q4 2008. It has been showed in many roadmaps that performance chips are also coming in Q4.

Lastest Intel Roadmap.

Metroid.

[gojirasan]

After every tick however, comes a tock.

Damn. Nice benchmark :banana::banana::banana::banana:. Gives me the chills. Cue The Red Dress from The Matrix soundtrack or maybe Meltdown from The Insider soundtrack. I get that "aliens have just landed on your front lawn" feeling. I keep thinking about how seriously badass Intel has become. Born again hard. Too bad most of my apps are single threaded. I take it the gains won't be quite so dramatic for single threaded stuff. As for multicore being the future, not everyone agrees. Donald Knuth, author of The Art of Computer Programming for instance has some issues with the idea. Nevertheless Intel's badassness is indisputable at the moment.

[fiveprime]

It's already about 9 months old.

I disagree with most folks. Intel will launch a few None Server Nehalems, maybe the Top two models for the desktop.

2.66GHz one and 2.83 is shown on their Roadmaps for Q4-08.

Maturation as in time on the market, as in, see how it preforms.

Unless ofcourse Intel will make these 2P and 4P exclusively at launch, you should be able to get a Nehalie for $999 and a mobo for $250 so $1,250 is a much better estimate.

Perkam

1k for a cpu? No thanks, not with the 45nm C2Qs being so cheap, especially since when working with a single thread it only gets a small boost.

[WeStSiDePLaYa]

Anyone else notice single thread performance only up by 2%?

[Metroid]

Anyone else notice single thread performance only up by 2%?

No surprises here as the memory improved but issues lock it out, single threaded applications will benefit from the memory latency/bandwidth and minimum from the cache etc.

Hope you have had a happy birthday anyway

Metroid.

[fiveprime]

No surprises here as the memory improved but issues lock it out, single threaded applications will benefit from the memory latency/bandwidth and minimum from the cache etc.

Hope you have had a happy birthday anyway

Metroid.

That may be true, but will it be enough?

[Metroid]

That may be true, but will it be enough?

Yes, most of the programs use 2 or more cores, no really a reason to worry much about it. I mean benchmark programs like SupePI will soon use more than a core but that will be a feature and not a removal of it.

Metroid.

[drfedja]

In this http://www.anandtech.com/cpuchipsets...px?i=3216&p=11
review Penryn QX9450 scores 3297 points for Cinebench 10 1-CPU. This is exactly faster than Nehalem's 3015 points. :d
Multithread scaling are better than C2D's.