Arrandale versus current i7

I just read http://en.wikipedia.org/wiki/Arrandale_(microprocessor)

According to that chart, does this mean that, although Arrandale will have built-in gfx, it will be lower performance than current i7? With 1/2 the cache and much lower TDP, I'd expect the clock speed and performance to be less also.

 

as far as I know, yes its going to have less performance.
http://www.engadget.com/2009/11/12/intel-arrandale-chips-detailed-priced-and-dated/
As you can see, lower prices, lower clock speed(around 1.06ghz). I think they are going to replace core 2 dues...
Intel wants the whole market. so wants all the buyers. with all budgets From Atom to i9

 

Arrandale will be a dual core, quad thread processor. The current i7s, Clarksfields, are quad core, octa thread processors. Arrandale will be the mainsteam processor line, while Clarksfield was designed for the high end.

 

Those are ULV cpus. The mainstream one's will have good clock speed( 2.4 to 2.66ghz) going up to 3.33ghz in turbo mode.

but arrandale will have half the cores and cache.

 

Hmmm, so will they post a threat in terms of performance when compared to an i7?

And I'm correct in saying that i9s are compatible with i7 sockets right

 

The i9s coming to notebooks will most likely be the desktop versions which will only fit into the desktop X58 motherboard (I think only one notebook has this), I doubt there will be a mobile i9... the mainstream Arrandales will be faster than the Clarksfield only when using programs that are not optimized to take advantage of all the cores, due to their clock speed being higher. That being said, I don't think there will be much difference nor will there be many programs that will only be optimized for dual cores.

 

remember that the current crop of i3/5/7/9 processors are just the first 'tick' of Intels tick/tock product cycle. Unless you ABSOLUTELY need the bells and whistles of an i7, stay with Core2 until the product refresh hits late next year.

If past releases are any kind of indication, a product refresh will include a die shrink, power efficiency improvements, and thermal management improvements along with similar reworkings of the support chipset(s).

Also, Larrabee is significantly delayed and may not be released as a standalone product. Look for Larrabee features to be integrated in a near-future release of a upgraded GMA GPU.

 

Nope. Westmere is the refresh of Nehalem (aka tick). Core 2 Duo was already supposed to be replaced by 45nm dual core Auburndale, but it was canceled (due to the economy) and 32nm Westmere got pushed up. So now Westmere is here to stay, with a minor refresh in mid 2010 (adding Celeron/Pentium and some more Arrandales), and then it is all being replaced by the "tock" 32nm Sandy Bridge architecture.

 

Any word on whether that minor refresh next year will include true quad-core Arrandales?

 

Be careful not to get hung up on development code names. The improvements in the NEXT refresh are as I wrote regardless of what their names were or are. I'm going on the handouts from the last IDC, not website or magazine writeups.

Intel said that the initial release of iXXX processors earlier this year would be quickly replaced by the current releases and that has happened. The real product refresh will occur in late 2010. As with the recent (and solved) 40nm process problems at TSMC, Intels cpu releases will be driven by improvements in process technology.

 

In early 2011 when Sandy Bridge is released is when you will see 32nm Intel quad cores with integrated graphics in notebooks.

What? I'm not hung up on code names, I just don't understand what you are talking about.

Released this year are quad core 45nm Core i7's. They will supplement the release of Arrandale Core i3/i5/i7's in 2010. Then in the 2nd half of 2010, more Arrandales will be released and the quad cores may get shrunk to 32nm. And in 2011 all this is replaced by Sandy Bridge.

What I am saying is that Westmere is the product refresh of Nehalem, and do not expect a big die shrink or anything in late 2011. 32nm is here to stay until late 2011/early 2012 when Ivy Bridge shrinks Sandy Bridge to 22nm.

 

http://www.xbitlabs.com/news/video/...ext_Generation_Platform_to_Support_GPGPU.html

-No Larrabee IGP until Haswell, which is 2013. If you can call that near-future, sure.
-Arrandale will beat low end Clarksfield like 720QM in most apps, thanks to the clock speed advantage, multi-thread or not. The worst it'll do is be on par.

 

Don't agree with this until I see it, unless you have benchmarks to show it.

 

no Larrabee GPU until 2013. True

Larrabee features including hybrid gpu and cpu/gpu pairing are scheduled much sooner than that. Intel (and Nv and AMD/ATi) always add subsets of new technologies in pieces before the full-blown introduction.

 

By the math, i7-720QM gets to 2.4 GHz on 2 cores with max turbo-boost, and 2.8 GHz on 1 core with max turbo-boost. Highest end Arrandale (i7-620M) is listed at 2.67 GHz base, with 3.33 GHz turbo-boost (not mentioned if that's 1 or 2 cores), so yes, clock-wise, it looks like the highest end Arrandale (priced about $30 less than the i7-720QM, mind) beats the lowest end Clarksfield by 2-4 bins. With a margin that relatively low, however, I'm not sure it'll beat the Clarksfield in a heavily multi-threaded application.

 

As pointed out in my post above, the high end of Arrandale could/would be competitive with the low end of Clarksfield. Note that calling them i7 can be a bit of a misnomer; high end Arrandale is also called i7, along with the entire Clarksfield line. You might want to look at http://en.wikipedia.org/wiki/List_of_Intel_Core_i7_microprocessors for specs on Clarksfield, and http://en.wikipedia.org/wiki/List_of_future_Intel_microprocessors for numbers on Arrandale and i9.

Note that 6-core i9 is a desktop processor, and only fits in a LGA-1366 socket, so the only notebook that will take it is the Clevo D900F(Sager NP9280). Not even desktops with Lynnfield chips (LGA 1156) can take it, only the ones that can take the first generation Bloomfield chips will fit it.

 

See, Clarksfield is only at 1.6GHz(1.73GHz Turbo) with heavily multi-threaded apps.

The 3.33GHz clock in Turbo Mode is for single core. I'm guessing it'll do 2.93GHz with 2 cores.

 

Sure, but that's 8 threads on 4 cores, which means that even if Arrandale is running at 2.93 on 2 cores, that's only 4 threads on 2 cores, so on that heavily threaded application (ignoring overhead for core coordination) each Arrandale core will have to do twice the work of each Clarksfield core, so comparatively, the Arrandale will have its speed cut in half (or double the Clarksfield speed) and Clarksfield will beat out Arrandale (1.6/1.73 vs 1.47 if you cut Arrandale, 3.2/3.46 vs 2.93 if you double Clarksfield).

 

Clock speed is a uniform way to increase performance, no matter what you are running. It's not limited by interconnects, cache coherency or w/e, even software compatibility. But with cores, that's not true. Any application that fails to scale almost linearly will be slower on the much lower clocked, 720QM.

And another thing: http://www.xbitlabs.com/news/video/...ext_Generation_Platform_to_Support_GPGPU.html

Arrandale/Clarkdale IGP to support GPU encoding in later drivers.

 

But my whole point is that it's not really that much lower clocked when you compare apples to apples (dual and single core operation). Admittedly, if you toss in IGP GPGPU computing, Arrandale will pull further ahead, but we'll have to wait for numbers to see how much that actually helps.

 

I doubt if GPGPU computing will become mainstream in the next year.

 

?

65-70% is quite a bit of difference.

720QM only clocks at 2.4/2.8GHz for dual/single core operation, Arrandale 2.66GHz will do 2.93/3.33GHz for the same. So the rest of the comparison is up to heavily multi-threaded, ie 4 core.

 

How do you get 65-70%? I do admit that initially I thought Turbo might only be for single core, so my 2-4 bins was 2.4/2.8 vs 2.66/3.33, but if Arrandale can Turbo on 2 cores, it's just a straight 4 bins, pretty much. But even then, 0.5 GHz out of 2.4 GHz, is only about 20%, not 65-70%... Still, we're getting away from my original point. My original point was to dispute your "Arrandale beats Clarksfield even in heavily multi-threaded applications". The rest was just that even in non heavily threaded applications, it's not quite as far behind as the base clock speeds make it seem.

 

???????????

You must be really confused:

Elementary school math friend.

1.73GHz Clarksfield 4 cores vs. 2.93GHz Arrandale 2 cores, or 1.6GHz vs 2.66GHz.

2.93/1.73=69%

 

Except that you'll never get the Clarksfield on 4 cores running against the Arrandale on 2, unless you disable Hyper-Threading on the Clraksfield and leave it enabled on the Arrandale (which is hardly fair). If you have a program that makes the Clarksfield run 4 cores, the speed on the Arrandale will be (effectively) cut in half (since it'll have to double up the threads on its 2 cores). So, if it's 1.73 GHz Clarksfield on 4 cores, it'll be 1.47 GHz (effective) for the Arrandale on 2 cores (for the same load). If the load for the Arrandale only uses 2 cores, then the load on the Clarksfield should also be only 2 cores, so the Clarksfield won't be running at 1.73 GHz, it'll be running at 2.4 GHz. That's why I was mentioning apples to apples... any load that would only use 2 cores of Arrandale would also only use 2 cores of Clarksfield, so there's no reason it shouldn't shut down to 2 cores and go full turbo-boost for 2 cores (bar thermal issues, which we've been ignoring).

Now, I admit the numbers might get very confusing if we're running 5-7 threads, (3-4 cores on Clarksfield and 2 cores (overqueued) on Arrandale), so I haven't been trying to run those numbers.

 

Your knowledge is worse than I thought.

The above tells how much you know about CPUs. Not much.

To answer the worst part of the paragraph, which is the one I bolded, yes you will. Example is running on apps that run more than 4 threads. Clarksfield will then power up all 4 cores, while Arrandale will still only run 2(because it only has two if it wasn't obvious).

Windows 7 makes it even clearer: Windows 7 will allocate physical cores before using logical cores. Meaning exactly 4 threaded app will run 4 cores on Clarksfield and 2 cores + Hyperthreading on Arrandale. Perfectly fair.

The biggest point is of course this: The high base clock speed of Arrandale will allow it to defeat lower end Clarksfield(720QM) in most apps.

 

And you completely managed to miss the sentence after that bolded part. If you have something running 4 threads (assuming hyper-threading is off on both Arrandale and Clarksfield) then Clarksfield will power up all 4 cores, yes, and run 1 thread on each core, while Arrandale will have to share time with 2 threads on each core (without hyperthreading, that should make Arrandale run half as fast). If you leave hyper-threading enabled on both, then why wouldn't the Clarksfield just run 2 cores hyper-threaded to match the Arrandale, bar the possible Windows 7 issue? We are also, of course, assuming that a logical core is just as fast as a physical core at this point.

I'm not sure that Windows 7 will allocate physical cores before using logical cores. If I remember how thread parking works, it'll won't swap cores if that will result in a drop in performance, which means that if swapping from a logical core to a physical core will result in a drop in performance (because it'll have to copy over all the cache memory and queue) then it won't. This would seem to imply that there's some way for it to discriminate that operating 2 cores hyper-threaded may be better than operating 4 physical cores. Or, contrariwise, this may mean that the fact that the Arrandale needs to use hyper-threading means it's taking a hit in performance and thus can't actually be functioning at it's listed clock speed.

I'm not actually arguing that the i7-620M doesn't beat i7-720QM in most things except maybe highly threaded applications, I'm just arguing that the difference might not be nearly as high as it seems.

 

I'm just waiting for core i9 and ATI HD 5870 Mobile in my D900F....

 

^^^^^ non-related.


Ok, as to the question of whether Clarksfield will be refreshed, yes it will, but it is not (in theory) due until Sandy Bridge (new architecture).

Clarksfield and Arrandale are to live together. They make the Intel Mobile Nehalem line-up. Arrandale is to live alongside Clarksfield, being the mainstream CPU for all laptops using Intel, having the ULV, LV and the full versions. Clarksfield lives on another bracket. The upper side of Mobile Nehalem, it is a quad core, destined to heavy work and for top-of-the-line hardware only (this is excluding the hype for i7 that has spread to multimedia laptops that will eventually change to Arrandale).

Nehalem is the architecture and Clarksfield and Arrandale are the "Montevina" (reference to Conroe) of this platform. Whether there will be a small refresh of these, such as the "Montevina Plus" is something I am not aware of, possibly there will be one, but dont expect Clarksfield to be fully changed to redone.

Clarkdale is desktop, why was it even mentioned? Clarksfield is mobile. Arrandale is the mobile Clarkdale. And Clarksfield is the mobile Bloomfield. (I hate Intel naming).

As far as Arrandale beating Clarksfield. It is plausible, on single threaded apps, even on dual threaded apps. But Arrandale wont be able to match the 720QM on heavily threaded situations. Albeit the 4-cored behaviour of Arrandale, Clarksfield has a 8-cored one. It is slower, but don't judge by clock speeds, since that is useless nowadays. On programs that support multiple threads, the performance increase is almost linear as you add up cores.

Comparing the highest end i7d to the lowest i7q isnt the most accurate either, it is like comparing the T9900 to the Q9000. On some cases the T9900 will smoke the Q9000, on some others it is the other way around. My opinion is that this applies to i7 too. The top of the line i7d will outperfom the lowest i7q on some tasks, but the extra cores bear an edge when fully supported.

 

The original reason I posted my response is that just because 720QM will be priced higher and its a quad core, doesn't mean that it'll beat the 620M which costs a bit less(Intel's pricing looks about $50 difference) and is a dual core, due to the fact that the clock speed difference is vast.

I never said that the difference between 1 and 2 core performance will be large, just that it'll be higher on Arrandale(due to higher clock) and probably be just as fast, if not faster on highly multi-threaded apps(like 5-8 threads).

If Clarksfield has all 4 cores running, but not the logical cores, it'll definitely lose against 2 core which clocks 70% higher AND has the advantage of Hyperthreading.

Windows 7 behaves differently depending on the versions you have. Core Parking is used in server versions, and SMT parking is used in consumer versions. Core Parking allows logical cores to kick in before the physical cores(to lower power consumption), while SMT parking doesn't use logical cores until all physical cores have been turned on.

Before Windows 7
4-8 threads
1.73GHz Clarksfield vs. 2.93GHz Arrandale*

2-4 threads
2.4GHz Clarksfield vs. 2.93GHz Arrandale

1-2 threads
2.8GHz Clarksfield vs. 3.33GHz Arrandale

After Windows 7
3-8 threads
1.73GHz Clarksfield vs. 2.93GHz Arrandale*

2 threads
2.4GHz Clarksfield vs. 2.93GHz Arrandale

1 thread
2.8GHz Clarksfield vs. 3.33GHz Arrandale

*This is the comparison I'm talking about. It's obvious Arrandale will beat it in other modes.

 

But that's the thing. Clarksfield also has the advantage of Hyperthreading. So beyond the whole OS issue of core assignment, in a 4 thread application why wouldn't the hyperthreading and turbo-boost of the Clarksfield kick in and make it 2.4 GHz Clarksfield vs 2.93 Arrandale? That was the whole point of my apples to apples comment, since they both incorporate the same features, why couldn't/wouldn't the Clarksfield go to 2 cores to match the Arrandale? I do admit the bolded comment was poorly phrased, it probably should have read something like "For a loaded Arrandale, the Clarksfield can match it by using only 2 cores instead of going to 4 and dropping speed, unless you disable Hyperthreading".

As for whether or not Arrandale or Clarksfield will win when we get to 8 threads (best case for Clarksfield, worst case for Arrandale), I think it will, (since Arrandale will have to double up threads on each physical and logical core, so each core will have to do twice the work of the comparative Clarksfield core) and you think it won't. I don't think we'll be able to convince each other at this point, so let's just wait for someone to run benchmarks and we'll see the results then.

As for SMT parking, why use that instead of core parking unless hyper-threading involves some sort of hit in processing power/speed? This would seem to imply (making a lot of assumptions here) that 4 cores at 1.6 (using this one for easier math) are better than 2 hyperthreaded cores at 2.4, implying that using logical cores instead of physical cores is approximately a 50% penalty. If that is accurate, then a hyperthreading Arrandale at 2.93 GHz would "effectively" be at 2 GHz compared to physical cores, and with Clarksfield at 1.6 GHz... oh, we're back at about 25% boost again. Note that this paragraph is totally speculative, and the numbers are almost certainly wrong.

Really, I think my whole contention is that I highly disagree with you that Arrandale will be 65%-70% faster than Clarksfield across the board when it comes to actual performance.

 

I think I concluded that on my last post.

I don't really need to show you anything but whatever is below:

Here's a good comparison between cores and clocks:
http://ixbtlabs.com/articles3/cpu/archspeed-2009-6-p5.html
http://ixbtlabs.com/articles3/cpu/archspeed-2009-5-p9.html

51% performance increase with 64.5% frequency increase.
30.2% performance increase with 100% core count increase(from 2 to 4 cores).

 

I think I'm missing something here. What your first link seems to say is that when comparing 1.73 GHz to 2.93 GHz, the perfomance increase will not be the linear 69% performance increase, but something smaller, so something like the 51% you're mentioning there?

The second link seems to say, from comparing the first 2 charts, that hyperthreading imposes an approximately 16% penalty (probably more, actually, since the first chart is with both off, and the second chart is with both on, so the turboboosting is probably making the hyperthreading look better than it is).

So with those numbers, then in what is presumably best case for Arrandale versus Clarksfield (4 threads, Clarksfield running on 4 physical cores and no logical cores), then the Arrandale comes out at something like 43% faster... so where's the 65-70%?

 

1. Yes, sure, but 2 to 4 cores do worse than the only 64.5% clock speed increase.

2. I don't get where you get the 16% penalty with Hyperthreading here. I can't even see a benchmark with 16% loss, let alone entire page.

Seperate Hyperthreading-only tests show in average, it alone is responsible for 10% increase.

So much for Hyperthreading being regarded as a "loss"
http://ixbtlabs.com/articles3/cpu/archspeed-2009-3-p1.html

(Even then, my last post was purely about how 1-4 cores scale, no gain or loss with TB+HT will change the actual scaling numbers, well maybe a tinny bit)

3. Again, I think we have misunderstandings. Never claimed 65-70% advantage. Only point is that a lowly dual core that's $50 cheaper will outperform a supposedly higher end quad due to its high base clock.

 

1) The first link doesn't mention anything about cores, just that a higher clockspeed isn't linearly more effective, i.e. doubling your clockspeed does not generate double the performance. I think I ended up just tossing this point in to point out that clock speed isn't necessarily a uniform way to measure performance.

2) I was comparing the very first 2 charts in the second link. The first chart shows scores for 1/2/3/4 cores with HT and TB disabled. The second chart shows 2/4/6/8 cores with HT and TB enabled. I'm pretty sure the cores they list in the second chart are combinations of physical and logical cores. Thus I compared 2 physical cores in the first chart to 1 physical and 1 logical core (2 cores) in the second chart, as well as 4 physical cores from the first chart to 2 physical and 2 logical cores (4 cores) from the second chart to get a basic idea of how much worse a combination of physical and logical cores compares to purely physical cores. As I stated before, the fact that the second chart also includes turboboost might mean that the logical cores were performing at higher clockspeeds than they "should", but there's no way to measure that without more data that isn't in the chart.

My "penalty" was for how much penalty you might suffer for using a logical core instead of a physical core. Overall, yes, Hyperthreading will give a bonus, because you go from 4 cores to 8, even if those 4 added cores aren't quite as efficient as real, physical cores.

3)

 

Of course, nothing is.

However, doubling clock speed is more effective than doubling cores from 2 to 4 don't you agree? Even not doubling, only 64.5% increase beats 100% increase in cores overall.

Clock speed has the best scaling in terms of performance(6/10 to 8/10). Of course you can't double clock speed so easily.

I see where you get 65-70% from. That is true in heavily multi-threaded applications. The rest came because you assumed (that I said) that was true for every other case. But its the heavily multi-threaded apps where you'd EVER want to pay for the expensive, relatively power hungry 4 core processors with Hyperthreading.

It's STILL an overall net gain for Hyperthreading.

I see where you are getting the 16% number from:
Maya: 1 core=3.59
Maya: 1 core + HT=4.17

That's a 16% difference there. You forget higher score is a better score so Hyperthreading actually helps 16% here, not the other way around. See the up and down arrows next to the application names? That indicates which way you should be looking for comparison.

Apps with Up arrow=Higher score is better
Apps with Down arrow=Lower score is better

And you completely ignored this link which showed HT-only comparison: http://ixbtlabs.com/articles3/cpu/archspeed-2009-3-p1.html

 

No, I didn't ignore that, I just didn't feel that it brought anything productive or destructive to the argument that I was making (at the time). My point is a straight comparison of logical cores and physical cores. As I pointed out, "Overall, yes, Hyperthreading will give a bonus, because you go from 4 cores to 8, even if those 4 added cores aren't quite as efficient as real, physical cores." In your Maya example, it's really:

Maya: 1 (physical) core = 3.59
Maya: 1 (physical) core + HT (1 logical core) = 4.17 <-- Note that I consider this to be 2 cores!

So for your link, to my point of comparison of physical and logical cores, it doesn't help, because simply turning on HT doubles your "total" cores. My point wasn't that hyperthreading isn't an overall addition to the efficiency of the CPU, my point is that a "hyperthreaded" (logical) core isn't as efficient as a physical one, going back to the whole Arrandale vs Clarksfield argument we were having earlier.

Also, I thought we were largely talking about heavily multi-threaded applications, which is why I made the point for what is (presumably) the best case for Arrandale, which was 4 threads, with Arrandale running 2 physical and 2 logical cores (due to hyperthreading) and Clarksfield running 4 physical cores, and that the advantage there wouldn't be 65%-70%. I would think things would just get worse for Arrandale if you moved up from there with more threads than that...