Nehalem multithreaded performance/price: Apple's benchmarks

[Weepul]

Based on Apple's benchmarks (http://www.apple.com/macpro/performance.html), I think it's safe to approximate the 2x2.93 GHz Nehalem Mac Pro to be about 1.3x the speed of the 2x3.2GHx Harpertown for well-multithreaded applications that utilize all the multi-core performance available.

I think it's also safe to assume that performance between different speeds of the same family of processors can be approximated as being linear, eg. twice the clock speed is twice the performance.

So, let's do the numbers (considering only CPU performance, not GPU-accelerated processes)...

Using a 2x2.8 GHz Harpertown as the baseline (since it was the best performance value available in the previous generation, pricing source from EveryMac (http://www.everymac.com/systems/apple/mac_pro/stats/mac-pro-eight-core-2.8-2008-specs.html)), but adding US$137.95 for 4GB RAM (http://eshop.macsales.com/item/Other%20World%20Computing/64FB1MPK04GB) to get its memory equal to the Nehalem models (the Harpertowns came with only 2GB RAM):

2x2.8H (US$2937): 1x performance, 1x price, 1x performance/price
2x3.0H (US$3737): 1.07x performance, 1.27x price, 0.84x performance/price
2x3.2H (US$4537): 1.14x performance, 1.54x price, 0.74x performance/price
(relative to original price 2.8 GHz Harpertown Mac Pro upgraded to 6GB RAM from OWC)

Since Apple's benchmarks compare the two most expensive variants of the current and previous generations, the conversion goes like this:

1.3*(Nehalem speed/2.93)*(3.2/2.8)

2x2.26N (US$3299): 1.15x performance, 1.12x price, 1.03x performance/price
2x2.66N (US$4699): 1.35x performance, 1.60x price, 0.84x performance/price
2x2.93N (US$5899): 1.49x performance, 2.00x price, 0.75x performance/price
(relative to original price 2.8 GHz Harpertown Mac Pro upgraded to 6GB RAM from OWC)

Remarkably similar to the previous generation. With the base model, provided you make use of well-multithreaded applications, you're getting the performance of the previous 3.2 GHz machine without the performance/price penalty that the original 3.2 upgrade had.

So, why is this update causing such a dissatisfied stir? Well, because performance/price did not improve much at all, which is contrary to expectations of updated technology.

For comparison, when the 2x2.8 GHz Harpertown came out, it was approximately 1.8x as fast (http://www.macworld.com/article/131782/2008/01/macprobench2.html?t=205) as its 2xDual-core 2.66GHz predecessor (and that's being generous (http://www.barefeats.com/harper.html)) for 1.12x the price, or a 1.6x performance/price improvement. That's what's missing with the Nehalems.

Furthermore, if you can snag a 2x2.8 GHz Harpertown for nearly any discount, you're coming out ahead for your money, as long as you're satisfied with a little less absolute speed.

That's some great technological progress, huh? :(

Independent benchmarks should clarify this further.

Thoughts?

[Tesselator]

'...for well-multithreaded applications that utilize all the multi-core performance available."

Being the key phrase there?

Nice post BTW...

[More]

Nice post. Makes me feel even more content with my recently purchased 2008 Mac Pro 3.2Ghz. Plenty fast for everything I do these days. There's also something nice about having the fastest Mac Pro of the 2008 generation rather than having the slowest Mac Pro of the 2009 generation (price/performance).

[Gonk42]

Interesting comparison. In the US the deal is reasonable but not brilliant (given that much software won't be well multi-threaded).

In the UK (and Australia is similar) the price rise is enormous. I was looking at getting 16GB of RAM from Crucial with a student discounted 2.8 octo plus a few bits like apple care etc was £2000. The equivalent with the new 8 core is something like £2,600 so the ratio is 1.3 giving the price performance of just
0.88.

On top of that, if I want to keep to 3 channel memory and not spend a fortune on 4GB chips I am now limited to 12GB instead of 16GB.

It is not just the ratio, in absolute terms it is hard to find an extra £600 plus and if I'm to drop to 4 core I might as well just get a consumer machine which will be less RAM limited (though not ECC) and be cheaper.

[DoFoT9]

great post weepul!!

i know that has clarified a few things for me :)

just a question, for price/performance, its better for the number to be above 1.x isnt it??? (got myself confused).

[Weepul]

'...for well-multithreaded applications that utilize all the multi-core performance available."

Being the key phrase there?

Of course. While the performance of less-threaded apps (a good number of Photoshop operations, for example) is a good thing to keep in mind, people looking to get the 8-core machines, particularly when there's a 4-core model available*, ought already have a use for multithreaded performance.

*RAM limit notwithstanding...

great post weepul!!

i know that has clarified a few things for me :)

just a question, for price/performance, its better for the number to be above 1.x isnt it??? (got myself confused).
Thanks. :) The higher the ratio in my post, the better - since I used performance/price (higher is better seems logical to me :D). Price/performance as you wrote (and as I more commonly hear said) is better if lower.

[DoFoT9]

lalala


Thanks. :) The higher the ratio in my post, the better - since I used performance/price (higher is better seems logical to me :D). Price/performance as you wrote (and as I more commonly hear said) is better if lower.

aahh righto, its KIND (sort of) dissapointing to see the 0.x results, especially in apples top spec machine.. i guess with intel now they are just interested in getting things out as fast as possible and not giving people like apple a chance to introduce the products properly. i feel it was kind of forced upon them..

thanks for verifying, i should have checked my post before replying!!

[Weepul]

aahh righto, its KIND (sort of) dissapointing to see the 0.x results, especially in apples top spec machine..

Well, with the higher-performance options, you usually sacrifice some of the value for your money in order to get more speed in one machine. What's disappointing is that the base model is hardly a better value than the old one.

[DoFoT9]

Well, with the higher-performance options, you usually sacrifice some of the value for your money in order to get more speed in one machine. What's disappointing is that the base model is hardly a better value than the old one.

yes quite disappointing. i can see the go with the price because its only a recently introcuded chip, meaning it costs more.. but you know.. not fair !

[Weepul]

Hmm, when considering value, Tesselator did bring up a good point: a quick search revealed that Nehalem single-threaded performance (http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3326&p=7) is near equal to Penryn performance, so while the 2.26 GHz Mac Pro might equal the previous 3.2 GHz with a multithreaded app, Safari will run like it's a 2.26 GHz either way, unless I'm missing something or that article is out of date. :(

[DoFoT9]

Hmm, when considering value, Tessellator did bring up a good point: a quick search revealed that Nehalem single-threaded performance (http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3326&p=7) is near equal to Penryn performance, so while the 2.26 GHz Mac Pro might equal the previous 3.2 GHz with a multithreaded app, Safari will run like it's a 2.26 GHz either way, unless I'm missing something or that article is out of date. :(

hhmm thats a very good point you bring up there, what will happen to single-threaded applications. they would be GREATLY disadvantaged by that low end machine..

[robinp]

Hmm, when considering value, Tesselator did bring up a good point: a quick search revealed that Nehalem single-threaded performance (http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3326&p=7) is near equal to Penryn performance, so while the 2.26 GHz Mac Pro might equal the previous 3.2 GHz with a multithreaded app, Safari will run like it's a 2.26 GHz either way, unless I'm missing something or that article is out of date. :(

no that is my understanding too, which is why for most games, the i7 was not considered good value over the previous generation. As a result, I am trying to get a good deal on one of the old 3.2 harpertown mac pros.

[More]

no that is my understanding too, which is why for most games, the i7 was not considered good value over the previous generation. As a result, I am trying to get a good deal on one of the old 3.2 harpertown mac pros.

A really interesting point! Kind of reminds me of the Dual 2.7Ghz G5 compared to the Quad 2.5 G5. The 2.7 always had that extra 0.2Ghz of horsepower for single threaded apps like Illustrator but the Quad 2.5 always kicked ass on multi-core operations.

Most consumer level apps out there don't even touch additional cores. My most used app, Illustrator CS3/4 still barely uses more than one core. My 2008 Mac Pro 3.2Ghz barely breaks a sweat when using apps like this.

Basically it's all down the the quality of software and it's multi-core implementations. Most apps are wasted on the last generation's power never mind the crazy speeds of the new models.

For me, I'd much rather have the power of 3.2Ghz on a single core app than a measly 2.26Ghz unless anyone can confirm or prove that the new intel chips at that speed can kick the ass of a 3.2Ghz chip on a single core.

[tamvly]

great post weepul.

i agree with other posters on this one - despite the new architecture, what most people wanted from this new apple system was a continuation of improved price-performance we've all come to expect as the natural march of technology.

this is a real disappointment for new buyers, while current harpertown owners just got a big boost in the value - and longevity - of their systems

anticipating that i would finally upgrade to a new mac pro, this announcement is a good reason to stick with my 2x2.5 g5, as long-in-the-tooth as it might be. i'll wait it out until there are some significant price-performance improvements.

my sense is that apple will not generate great sales numbers for this system.

[Daim]

'...for well-multithreaded applications that utilize all the multi-core performance available."

Being the key phrase there?


that's there snow leopard comes in giving more and more applications the chance to use multi-core technology (+gpu) more efficiently

[Infrared]

Nehalem models (the Harpertowns came with only 2GB RAM):

2x2.8H (US$2937): 1x performance, 1x price, 1x performance/price
2x3.0H (US$3737): 1.07x performance, 1.27x price, 0.84x performance/price
2x3.2H (US$4537): 1.14x performance, 1.54x price, 0.74x performance/price
(relative to original price 2.8 GHz Harpertown Mac Pro upgraded to 6GB RAM from OWC)

1.3*(Nehalem speed/2.93)*(3.2/2.8)

2x2.26N (US$3299): 1.15x performance, 1.12x price, 1.03x performance/price
2x2.66N (US$4699): 1.35x performance, 1.60x price, 0.84x performance/price
2x2.93N (US$5899): 1.49x performance, 2.00x price, 0.75x performance/price
(relative to original price 2.8 GHz Harpertown Mac Pro upgraded to 6GB RAM from OWC)

Very interesting post, thanks. I've reworked your figures for the UK:

Nehalem models (the Harpertowns came with only 2GB RAM):

2x2.8H (UK£1712): 1x performance, 1x price, 1x performance/price
2x3.0H (UK£2202): 1.07x performance, 1.29x price, 0.83x performance/price
2x3.2H (UK£2701): 1.14x performance, 1.58x price, 0.72x performance/price
(relative to original price 2.8 GHz Harpertown Mac Pro upgraded to 6GB RAM from OWC)

1.3*(Nehalem speed/2.93)*(3.2/2.8)

2x2.26N (UK£2499): 1.15x performance, 1.46x price, 0.79x performance/price
2x2.66N (UK£3619): 1.35x performance, 2.11x price, 0.64x performance/price
2x2.93N (UK£4579): 1.49x performance, 2.67x price, 0.56x performance/price
(relative to original price 2.8 GHz Harpertown Mac Pro upgraded to 6GB RAM from OWC)

[Infrared]

that's there snow leopard comes in giving more and more applications the chance to use multi-core technology (+gpu) more efficiently

Care to share your Snow Leopard benchmarks with us?

[Dysfnctnl85]

Care to share your Snow Leopard benchmarks with us?

Wouldn't this only affect applications running OpenCL code?

[nicolasmasset]

Yeah, I thought this whole grand central thing and snow leopard were going to make single core processes multicore?

[mason.kramer]

Yeah, I thought this whole grand central thing and snow leopard were going to make single core processes multicore?

No. That's not possible. To the extent that an application can be multithreaded automatically, it is a function of the compiler used to compile it (e.g. auto loop vectorization). To the extent that a single thread of execution can be executed out of order (OOE), it is a function of the CPU microcompiler.

The OS doesn't have much to do with auto-multithreading. Developers have to build multithreaded applications. You know when Apple says "all applications will be multithread aware?" Yea. That's either a lie, or a meaningless statement designed to mislead.

[nicolasmasset]

No. That's not possible. To the extent that an application can be multithreaded automatically, it is a function of the compiler used to compile it (e.g. auto loop vectorization). To the extent that a single thread of execution can be executed out of order (OOE), it is a function of the CPU microcompiler.

The OS doesn't have much to do with auto-multithreading. Developers have to build multithreaded applications. You know when Apple says "all applications will be multithread aware?" Yea. That's either a lie, or a meaningless statement designed to mislead.

Hm, so a 2008 2x2.8 machine will be faster than a 2900 2x2.26 machine for single thread apps? Well actually, the only "light" apps I use that I presume are single thread are safari and mail and occasionally dreamweaver. The only "heavy duty" apps are ableton live and logic with a plethora of plugins.

So I guess my first argument here isn't really relevant. I guess the proper question would be: do ableton and logic and reverb plugins etc. use multiple cores efficiently to be able to benefit the new macs?

[Dr.Pants]

I guess the proper question would be: do ableton and logic and reverb plugins etc. use multiple cores efficiently to be able to benefit the new macs?

I'm also interested in this and (unfortunatly, since all of my posts mention FCP) also have to beg the question; when it comes to useability in Final Cut (not the render, just using the program), would a better investment be in an upgraded 2.6 gHz Quad from the 2.26 gHz octo?

I mean, I spend about eight hours a day doing university-stuff, so I think she'll have plenty of time to do some renders... I'm just concerned about the actual usage of the program.

However, another thing I am concerned about is capturing 1080 progressive to RAID.... (30 FPS, for those who are interested). Would a 2.26 octo or 2.66 quad system be better for it?

[skyline r34]

This why i'm not buying the 2009 Mac Pro, I have a 3.2GHz and in what I seen on Apple website about the 2009 Mac Pro's it's a small performance gain, i'm just gonna upgrade my video card to the ATi 4870 so that I can use the LED 30 inch when it's release and that's it and wait until the next Mac Pro next year 2010 hope by then we will see some 16 cores and better choices of Gpu's, Hope Westmere will have better numbers but the one good thing about the 2009 Mac Pro is the multithreaded but other than that i'm not to impress by the new Mac Pro.

[Quash]

I read a lot of these topics. And i see the point of em.

But i don't see why the angry reactions towards Apple. Unless your still on about the "x-mac" with desktop CPU.... People have been shouting for it as long as i can remember. So i will say it one more time, repeat after me: As long as Steve Jobs has a controlling say in Apple it is never going to happen.

Steve sees computers as appliances. So a fully desktop based user serviceable mac doesn't fit that vision. So we can talk about it all day and compare Xeons to Normal Desktop Core i7 systems. Still not going to happen.

Anyway the thing is though apple is now stuck to Intels roadmap. Intel publishes it's roadmap well ahead. Goes for either Xeons or mobile processors.

You can know what new processor line will be coming months ahead. This includes prices of the processors well in advance of release. More then a month ago these came out for the i7 Xeons and you could have seen then what you see now. (Which i posted in the rumor thread back then ;) ) They are much much more expensive clock for clock as the harper core 2 duos.

Which sucks but it isn't exactly apple's fault now is it. I doubt they changed their profit margins on these machines.

Same goes for the failing UK pound. It's just not strong these days, because of your economy. Meaning all imported goods will become much more expensive. Most product you don't notice this as it goes gradually as stock runs out. Apple only adjusts prices at new releases, doesn't change the fact their prices are coupled to the dollar (as all pc parts are btw).

[mason.kramer]

Hm, so a 2008 2x2.8 machine will be faster than a 2900 2x2.26 machine for single thread apps?

Nehalem architecture is faster, clock for clock, than Harpertown (&C2D) before it. Before going on with this, make sure you know what processor architectures you're comparing (2008 vs 2009 doesn't guarantee anything, but I can tell you that 2008 MP's offer Harpertowns and 2009 MP's offer Nehalems, but those are quad core parts, not dual core parts like you're implying).

Now, you can't compare Harpertown & Nehalem :apple: to :apple:: probably the Nehalem 2.26 will be faster in serial execution than the older generation 2.8, but no one knows for sure. Nehalem has
1) A wider execution pipeline
2) A more advanced instruction set
3) Higher memory bandwidth
4) Lower memory latency
5) The ability to increase its clock speed when other cores on the die are not dissipating power

All of these things contribute to make Nehalems faster in serial execution clock for clock. They also have hyperthreading, which makes them faster for multithreaded apps.



Well actually, the only "light" apps I use that I presume are single thread are safari and mail and occasionally dreamweaver. do ableton and logic and reverb plugins etc. use multiple cores efficiently to be able to benefit the new macs?
Logic does, I don't know about ableton. Safari and Mail are also multithreaded.

[Fomaphone]

I'm also interested in this and (unfortunatly, since all of my posts mention FCP) also have to beg the question; when it comes to useability in Final Cut (not the render, just using the program), would a better investment be in an upgraded 2.6 gHz Quad from the 2.26 gHz octo?

I mean, I spend about eight hours a day doing university-stuff, so I think she'll have plenty of time to do some renders... I'm just concerned about the actual usage of the program.

However, another thing I am concerned about is capturing 1080 progressive to RAID.... (30 FPS, for those who are interested). Would a 2.26 octo or 2.66 quad system be better for it?

don't forget that the next FCP will be rewritten to maximize this architecture and Snow Leopard.

also, go with the octo so that you can go beyond 8gb when the time comes.

[junnny]

Hmm, when considering value, Tesselator did bring up a good point: a quick search revealed that Nehalem single-threaded performance (http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3326&p=7) is near equal to Penryn performance, so while the 2.26 GHz Mac Pro might equal the previous 3.2 GHz with a multithreaded app, Safari will run like it's a 2.26 GHz either way, unless I'm missing something or that article is out of date. :(
That said, no one is talking about the single processor "Nehalem" option ... the 2.93GHz Quad-Core Intel Xeon ... which for an additional $500 (oops) has a good clock rate, four cores, and "Nehalem" architecture.

Maybe this is worth considering. Can anyone comment on the "pros & cons" of this choice?

[Nik]

That said, no one is talking about the single processor "Nehalem" option ... the 2.93GHz Quad-Core Intel Xeon ... which for an additional $500 (oops) has a good clock rate, four cores, and "Nehalem" architecture.

Maybe this is worth considering. Can anyone comment on the "pros & cons" of this choice?

Well, the 4-core config is a really good choice for using applications which do not support more than four cores at once.
Photoshop is such an app and will run about 30% faster on a 4x2.93GHz MacPro than it does on a 8x2.26GHz MacPro.
Unless you are rendering movies or CGIs the MacPro with four cores is a better choice.

Lets hope that the quad-core MacPro supports more than 8GB of ram.
In my opinion the 8GB cap is just a marketing gag. As soon as 4GB modules are available you can upgrade the machine to get 16Gigs of ram.

The performance of a 2.93GHz equals the performance of a 3.8GHz Xeon of the previous generation if a gain of 30% is realistic.

[bvk]

Thank you Nik - that's exactly what I was wondering as well. :)

[nanofrog]

That said, no one is talking about the single processor "Nehalem" option ... the 2.93GHz Quad-Core Intel Xeon ... which for an additional $500 (oops) has a good clock rate, four cores, and "Nehalem" architecture.

Maybe this is worth considering. Can anyone comment on the "pros & cons" of this choice?
Nik made some good points, and it would be a really nice machine. :D

The biggest downside is cost. The Quad cores use W35xx parts, (W3540 in this case), and is just an i7-940 with ECC enabled. Intel sells them for the same price as their Core i7 counterparts (respective of clock speeds).

Memory capacity may be another, as it can only handle 8GB (4x 2GB DIMM's). You'd have to locate and change out the stock memory and replace it with 4GB sticks, for a max of 16GB. Not an ideal situation. :(

[iMacmatician]

Based on Apple's benchmarks (http://www.apple.com/macpro/performance.html), I think it's safe to approximate the 2x2.93 GHz Nehalem Mac Pro to be about 1.3x the speed of the 2x3.2GHx Harpertown for well-multithreaded applications that utilize all the multi-core performance available.Great post and analysis. Also…

2.4| x
2.3|
2.2|
2.1|
2.0|
1.9| x
1.8| x
1.7| x
1.6| x
1.5| x x
1.4| x x
1.3| x x x
1.2| x x x x x
1.1| x

Median = 1.35x (1.3x excluding synthetic benchmarks)

That's some great technological progress, huh? :(Absolutely. :(

There's also something nice about having the fastest Mac Pro of the 2008 generation rather than having the slowest Mac Pro of the 2009 generation (price/performance).That's kinda like having an entry-level MacBook Pro instead of a high-end MacBook (maybe of the next generation). It's interesting to see what features the MacBook Pro has that the MacBook doesn't, and vice versa.

[mason.kramer]

Apple doesn't have benchmarks. They have graphs.

Just wait for somebody to post some benchmarks before we do this.

[nicolasmasset]

Nehalem architecture is faster, clock for clock, than Harpertown (&C2D) before it.

Logic does, I don't know about ableton. Safari and Mail are also multithreaded.

Thanks mason, that was a useful post!

[Abidubi]

The performance of a 2.93GHz equals the performance of a 3.8GHz Xeon of the previous generation if a gain of 30% is realistic.

Wouldn't a 2.93 equal a 4.16 of the previous generation if it really is 30% faster on average than the last 3.2?

1.3 X 3.2 = 4.16

[mason.kramer]

Wouldn't a 2.93 equal a 4.16 of the previous generation if it really is 30% faster on average than the last 3.2?

1.3 X 3.2 = 4.16

This is meaningless.

[Abidubi]

This is meaningless.

How is that meaningless? Thats like saying comparing the combined 10.64Ghz of the quad 2.66 vs the combined 18.08Ghz of the octo 2.26 is meaningless. It is very much meaningful because thats exactly what it works out to if you are going to be running something that will fully use every last core your machine has.

And thats why anyone doing that kind of work with their system wont even consider the 4 core version. I on the other hand, wont be using many programs that can exploit 8 cores.

Relative to the previous processor performance, the new top of the line does 33.28GHz of processing compared to the 25.6GHz of processing (again on the old scale).

[junnny]

Well, this thread sure hit a wall :) I guess everyone is waiting for independent benchmarks to help them make decisions, which makes sense. Oddly, there's some sort of performance/price thing that's been set up with the introduction of "Nehalem" ... particularly (at least for me) this 8-core 2.26ghz wrench that's been thrown into the works when compared to having a single 4-core 2.66ghz processor, never mind, a single 2.93ghz processor ... these latter machines which can be purchased for a lot less money than the higher end 8-core models, but even the 8-core 2.26ghz.

None of this is confusing if you're not planning to upgrade, don't think the power increase is that significant, upgrade every year or two, or can easily afford the highest end offerings, even if you don't need them. I'm planning to upgrade around June, though, with the idea that the release of the next operating system will optimize all this multi-core stuff ... including the single 4-core chip, which, let's not forget, is "multi-core." I'm not a computer jock, like some of you here, but I imagine this will help bring some of these issues into focus.

That said, and until the benchmarks arrive, I've read a little about how the Nehalem chip detects when it's not using some of its cores and boosts the processing speed of the remaining ones or one ... making what was a 2.66ghz processor behave like a, let's say 3.xghz processor. Does anyone know what I'm referencing and if so, how this would be applied to the 2.26 8-core, in particular? I must say, it's discouraging to pay so much for what is nominally (and for some applications) a 2.26ghz machine.

Thanks.

[Gonk42]

That said, and until the benchmarks arrive, I've read a little about how the Nehalem chip detects when it's not using some of its cores and boosts the processing speed of the remaining ones or one ... making what was a 2.66ghz processor behave like a, let's say 3.xghz processor. Does anyone know what I'm referencing and if so, how this would be applied to the 2.26 8-core, in particular? I must say, it's discouraging to pay so much for what is nominally (and for some applications) a 2.26ghz machine.

Thanks.

Here's a link to the Intel page on it (it's under core i7 at present as the
Xeon chips have not yet officially been released but it is the same for
both I understand).

http://www.intel.com/technology/turboboost/index.htm

My understanding from the white paper and from some other reading is that the clock can increase in increments of 133MHz and I guess it is something like 1 step for 3 cores active (out of 4) up to 3 steps for 1 core active (3 inactive) so the maximum frequency of a 2.26GHz chip will rise only to that of a 2.66GHz chip (it doesn't help that the total cores for 2 chips are 8 because it is the number of cores on the single chip that has the active core that matters as it is all to do with local heat generation/dissipation). The 2.93GHz chip can thus go up to about 3.33GHz.

I hope that this helps (and is right!)

[Weepul]

I sure hope Barefeats gets their system(s?) soon. :eek:

[Tesselator]

that's there snow leopard comes in giving more and more applications the chance to use multi-core technology (+gpu) more efficiently

<Rubs hands together vivaciously> <drools>
I sure hope so!!! :)