The Mac Hardware Report: A Look at the Eight-Core Value Equation
April 11th, 2007Over the next few weeks, you’ll be seeing lots and lots of benchmarks for Apple’s eight-core, 3GHz Mac Pro. Perhaps the fundamental question to be answered is whether it’s really worth paying a $1,498 surcharge over a pair of dual-core 2GHz Intel Xeons to get one.
Call this a preliminary and largely cursory analysis, based on very little external evidence, but my gut feeling is that it’s not.
Am I serious? Indeed, I am. You see, there has to be an incredibly large performance improvement to justify paying such a princely sum for a beefier processor pair. And there’s the rub.
Take, for example, the preliminary tests Rob-Art Morgan has posted over at Bare Feats, a Web site devoted to heavy-duty benchmarking. In some respects, the results are predictable. Both Cinebench and GeekBench show typically large performance increases when you go from four-cores to eight. However, the improvement is extremely modest — make that hardly noticeable — with the Public Beta version of Adobe Photoshop CS3 and the latest edition of Apple’s Aperture.
I suppose this mixed bag of results might be be highly disappointing for graphic artists, who were no doubt hoping that Photoshop would deliver blazingly fast performance, more than sufficient to justify the migration from four to eight cores. No such luck, although content creators specializing in 3D work will be delighted.
So why the disparity? Well, Rob-Art quotes long-time Mac programmer Lloyd Chambers (one of the creators of the legendary DiskDoubler years ago) as blaming the eight-core Mac Pro’s performance limitations on an overtaxed memory bus. In other words, the RAM isn’t capable of keeping up with the capability of the processors when they’re operating at fill tilt.
In addition, few applications can even handle two processor cores, so even going to four-core may be a waste of time and money. In fact, I suspect most of you would be perfectly satisfied with the standard-issue 2.66GHz Mac Pro, since the performance boost going to a pair of dual-core 3GHz Xeons would be far more modest than the price price increase.
Going Quad-core may be a terrific idea if you crave bragging rights, of course, but the expense would be seldom justified, except for that small subset of applications that can take advantage of all that mass CPU horsepower.
So why did Apple release such an upgrade in the first place?
Well, for one thing, I’m sure that the demand was there, created when Intel’s first four-core “Clovertown” processors first appeared last fall. However, that first batch topped out at 2.66GHz, which would actually be slower than the 3GHz dual-cores in many instances.
The 3GHz version of Clovertown is apparently the result of a limited production run that Intel designed for Apple. The likes of Dell and HP won’t get their hands on such processors until later, although it may not matter all that much.
Clearly Apple hasn’t made a huge deal of it either. You see, the eight-core Mac Pro option was only introduced with a press release. As of the time this story was written, Apple had yet to post benchmarks for an eight-core system.
In an interview for this week’s episode of The Tech Night Owl LIVE, Macworld’s Editorial Director, Jason Snell, said Apple seemed pretty low-key about the potential for the new Mac Pro with them as well. Apple suggested it would be ideal for programmers, in fact, who are developing applications that will take advantage of multi-core processors.
I also wouldn’t be surprised to hear that Leopard is being optimized for four cores or even additional cores. By next year, such processors will be more and more commonplace, and will likely appear in note-book computers as well.
So it may well be that the eight-core Mac Pro is, right now, simply an expensive vision of the future. If you’re up to spending a bundle to acquire one of these beasts, you’ll have a personal computer with a huge amount of potential, but you may also be chasing a rainbow with little hope of finding its end.
Print This Post
“I also wouldn’t be surprised to hear that Leopard is being optimized for four cores or even additional cores. ”
OS X isn’t optimized for a particular number of CPU cores. The kernel and the frameworks take advantage of multithreading, and those threads are scheduled across as many CPUs as you’ve got. The benefit you get will vary according to many different factors, with the main one being how well application developers implement multithreaded apps.
-jcr
They can always improve multithreading, which — sans the programming descriptions — is what I was getting at. Too bad Apple isn’t eating its own dog food, such as it is, particularly with the Finder.
Peace,
Gene
I just checked the Activity Monitor on my Quad G5, and the Finder has five threads when it’s idle. It jumped up to nine threads when I started moving things around, then up to 23 threads when I played a movie in the preview pane, and dropped on back to five when I closed some of the file viewer windows I’d opened. I’ve got my own gripes about the Finder, but I can’t complain about it ignoring threading opportunities.
At any rate, when it comes to the Finder, Apple is indeed eating its own dog food. What’s absurd is eating their own dog food (Carbon) when they could be eating their own filet mignon (Cocoa). The political reasons for the Finder being a Carbon app are now moot; hopefully it will be replaced in 10.6.
-jcr
In answer to what jcr wrote, here’s another quote from the same Bare Feats article: “the problem is that the OS does not know that it should keep a thread on the SAME core instead of swapping it around to any one of the 8 cores. When a thread moves to a different core-group, the cache has to be reloaded on the new core.” This is written about Tiger. This does indeed look like an area in which Leopard’s handling of multicore processing might be improved.
There are certainly issues with OS optimisation for threads. The OS has to perform functions for lots of applications, and sometimes two applications (or threads from the same application) are going to want to access the same parts of the OS. When this happens, locks have to be acquired and released. According to Ars Technica, Apple has been progressively making the kernel space more fine-grained with time, which in turn improves threading. Initially, there was just one big lock, which meant that only one application could call into the kernel at a time. These days, more than one application can be in kernel space at a time.
Why not do this from the start? First, concurrency issues can be very subtle, so getting things right is decidedly non-trivial. Secondly, all those locks impose an extra overhead, so you need to be sure that there’s enough work and enough places to put it for it to be worthwhile.
All this said, I would not be at all surprised if the major problem was with the memory bus – back to the old problem with the G4. With 8 cores looking for data, NUMA systems must be beginning to look attractive (provided Apple can also fix the core-swapping issue).
I plan to upgrade my work computer this August, and I plan to get the 8-core model knowing that the performance may not excel. But I won’t be upgrading again for another three years, so I am that case that is mentionedâ€â€not for bragging rights, but for “future-proofing” my computer.
[Comment ID #6165 Will Be Quoted Here]
Indeed, as a vision for the future, it may well make sense to buy the 8-core model. On the other hand, it’s also possible a revised version will come along with a chip that supports a speedier memory bus, so that other performance bottlenecks can be addressed.
But if you want the fastest Mac on Earth now, you have to realize that status will always change in a few months, regardless of where you begin.
Peace,
Gene
That’s why I never get the top-of-the-line at home. Then I’d be in a never-ending battle to keep up. It doesn’t hurt as much to get knocked off if you aren’t falling from the highest point. 😉
Those who need 8 core will buy 8 core: Creative folks working with audio and Video Codeces, people doing 3 d rendering, scientific/mathmatical, etc. Even compiling code would benefit greatly.