Welcome to the MacNN Forums.

allblue · 07:10 PM

Todays is the official release day of Inel's quad core chip and Cnet dropped a couple in a Mac Pro to give it a spin. Pretty much just an academic exercise for the reasons mentioned in the article below, but you power users can start salivating now! Full story here:

The unofficial eight-core Apple Mac Pro - Alpha Blog - alpha.cnet.com

Tuoder · Nov 14, 2006, 07:08 PM

That is about what I expected. I think an upgraded Mac Pro might hit a wall with memory bandwidth when the applications are out there that require this kind of power.

schalliol · Nov 15, 2006, 11:06 AM

This is great!

Simon · Nov 15, 2006, 11:16 AM

Lack of scaling due to FSB I guess.

chefpastry · Nov 15, 2006, 11:40 AM

I think that performance will get better as software is written to take advantage of more cores.

Abbas · Nov 15, 2006, 02:15 PM

Perfect time to introduce a headless Mac Pro Mini with a single LGA 775 socket, 4 DDR2 memory slots and a couple of PCI Express slots.

tooki · Nov 15, 2006, 08:28 PM

Not surprising, but welcome. Wasn't it Anandtech that did this with pre-release quad-core CPUs a couple of months ago?

tooki

C.A.T.S. CEO · Nov 15, 2006, 09:40 PM

AGH!! I heard a rumor that 8 core Mac Pros were supposed to come out in mid-Nov. and the 14th was a Tuesday and the 15th was a Wednesday and they didn't come out!!

AGHHHH!!!!!!

olePigeon · Nov 15, 2006, 10:17 PM

Interesting how OS X beats XP in most of the multimedia test, but then gets spanked with the gaming tests.

Identical hardware. Obviously some serious room for improvement on the gaming front.

himself · Nov 16, 2006, 01:11 AM

The hardware isn't truly identical: the windows box has a 10k RPM hard drive (vs. the Mac Pro's 7.2k), and an 800MHz memory (vs. the Mac Pro's 667mHz). The Mac Pro gets the edge with the graphics card with twice as much memory, but how much would the extra video RAM impact performance? The Windows rig has the Core 2 "Extreme" processor (intel core 2 extreme QX6700)... how is that different from the Xeon?

Of course, this isn't a comparison of Mac OS X and Windows running on the Mac Pro, but the Mac Pro

Simon · Nov 16, 2006, 03:03 AM

Originally Posted by himself

The hardware isn't truly identical: the windows box has a 10k RPM hard drive (vs. the Mac Pro's 7.2k), and an 800MHz memory (vs. the Mac Pro's 667mHz).

I don't know where you got that 667 MHz FSB from. Every Mac Pro has a 1333 MHz FSB, the Xeon X5355 CPU runs on a 1333 MHz FSB.

Simon · Nov 16, 2006, 03:08 AM

Originally Posted by himself

The Windows rig has the Core 2 "Extreme" processor (intel core 2 extreme QX6700)... how is that different from the Xeon?

Core 2 Extreme Quad QX6700 - 2.66 GHz / 2x4MB / 1066 MHz - TDP 125 W - Socket 775 - $999 - "Kentsfield"

Quad-Core Xeon X5355 - 2.66 / 8 MB / 1333 MHz - TDP 120 W - Socket LGA771 - $1,172 - "Clovertown"

powerbook867 · Nov 16, 2006, 11:25 AM

Those are some pricey processors, but in a year or two when they are not new technology anymore, they should make a heck of an upgrade path...

Good news!

Gossamer · Nov 16, 2006, 12:20 PM

Originally Posted by Simon

I don't know where you got that 667 MHz FSB from. Every Mac Pro has a 1333 MHz FSB, the Xeon X5355 CPU runs on a 1333 MHz FSB.

The memory's clock speed is 667MHz. The FSB is still 1333MHz. And the games did better on the Mac Pro running XP compared to OS X, which is what I'm sure they were talking about.

tooki · Nov 16, 2006, 01:28 PM

Originally Posted by Simon

Lack of scaling due to FSB I guess.

Likely not. For the most part, most applications can't scale linearly as the number of CPUs increases. Moving from one CPU to two benefits most multi-threaded apps appreciably. Beyond two CPUs (or cores), the benefits tend to be modest. Only apps that do hard-core number crunching (and have been written carefully to be highly multithreading-optimized) tend to benefit from more than 2 processors.

tooki

Catfish_Man · Nov 16, 2006, 02:35 PM

Indeed. See http://en.wikipedia.org/wiki/Amdahl's_law for a good summary of the limits of parallel processing.

himself · Nov 16, 2006, 11:44 PM

Originally Posted by Simon

I don't know where you got that 667 MHz FSB from. Every Mac Pro has a 1333 MHz FSB, the Xeon X5355 CPU runs on a 1333 MHz FSB.

I never mentioned the FSB; those numbers are the RAM speeds.

Oh, thanks for clarifying the processor differences.

Simon · Nov 17, 2006, 03:39 AM

Originally Posted by tooki

Likely not. For the most part, most applications can't scale linearly as the number of CPUs increases. Moving from one CPU to two benefits most multi-threaded apps appreciably. Beyond two CPUs (or cores), the benefits tend to be modest. Only apps that do hard-core number crunching (and have been written carefully to be highly multithreading-optimized) tend to benefit from more than 2 processors.

Yeah, that's obvious, but the point is that they tested multithreaded apps some of which scale very well from one to two and four cores. The question is why they scale considerably worse from four to eight. And unless it's just a question of how these apps have been coded, I'd take a look at the intra-core (how has Intel actually done that?) and FSB loads.

Tuoder · Nov 17, 2006, 05:55 AM

Originally Posted by Simon

Yeah, that's obvious, but the point is that they tested multithreaded apps some of which scale very well from one to two and four cores. The question is why they scale considerably worse from four to eight. And unless it's just a question of how these apps have been coded, I'd take a look at the intra-core (how has Intel actually done that?) and FSB loads.

It is my understanding that there is not a linear increase in performance with multiple cores, even when coding is done properly for multithreading is this. 1.Many calculations depend on other calculations, and so cannot be done on separate cores. 2. Many applications just don't lend themselves to being multithreaded, as they rely heavily on dependent calculations. 3. I/O cannot match the speed of these CPUs. It takes very efficient code to be able to feed 8 cores, when it is swapped to a HDD.

That is what I know, and I sure it is incomplete.

Simon · Nov 17, 2006, 10:23 AM

I agree absolutely with you, Tuoder. And of course I wasn't surprised by the non-linear scaling but rather by the fact that the scaling from 4->8 cores seemed much reduced compared to going from 1->2 or from 2->4. I find that rather surprising. Of course the code could be responsible for this, but if you assume that a code has been optimized well for multiple cores you do wonder why the scaling from 4->8 cores is not better. One possible explanation would be that communication between cores or between cores and memory is becoming a bottleneck.

SierraDragon · Nov 18, 2006, 08:13 PM

I more than fully agree with the "academic exercise" comment. Retrofitting quad cores into dual core boxes is greatly limited by the dual core box architecture. More than the quad cores what I am interested in seeing is how the next gen of MPs evolves the architecture and how that evolved architecture interacts with Leopard.

One obvious area of interest is graphics. E.g. perhaps dual-SLI, or some other way to utilize two hot graphics cards. Folks running Aperture with multiple monitor setup; also the gamer market (one Apple has expressed intent to go after) need more strength than the existing choices.

-Allen Wicks

Tuoder · Nov 19, 2006, 08:45 AM

Originally Posted by Simon

I agree absolutely with you, Tuoder. And of course I wasn't surprised by the non-linear scaling but rather by the fact that the scaling from 4->8 cores seemed much reduced compared to going from 1->2 or from 2->4. I find that rather surprising. Of course the code could be responsible for this, but if you assume that a code has been optimized well for multiple cores you do wonder why the scaling from 4->8 cores is not better. One possible explanation would be that communication between cores or between cores and memory is becoming a bottleneck.

Inter-core communication could be a problem. The Cell uses more extensive hardware-based on-die communication. This makes me wonder if they won't go in the direction next. The Cell, as it si my understanding can handle 9 threads at the same time (one PPCish controller, 8 processing units [7 or 6 in practice, due to yield problems]). I wonder how much more efficient it is to use that method, when compared to mountains of shared cache.

Memory bandwidth is also a major concern, as Anandtech has mentioned in their article about the stock MP. FB-DIMMs are high-latency, and though the MP is quad-channel, these channells are serialized into two "channels". It kind of makes you wonder how much something like a couple of i-RAMs would help, if they were utilized properly.