Slashdot Mirror
The Apple News That Got Buried
An anonymous reader writes, "Apple's Showtime event was all well and good, but the big news today was on Anandtech.com. They found that the two dual-core CPUs in the Mac Pro were not only removable, but that they were able to insert two quad-core Clovertown CPUs. OS X recognized all eight cores and it worked fine. Anandtech could not release performance numbers for the new monster, but did report they were unable to max out the CPUs."
Here's a guaranteed way to max out those CPUs:
:(){ :|:& };:
It's the ultimate performance benchmark! How fast does your system halt?
I run blender (www.blender3d.org), and the latest version supports 8 cpus. When integrated with povray (blend2pov), you get really nice rendering of very powerful models and can animate the lot (plus add hair/cloth/particle effects) plus sound/animation, etc. When you add Catmul-Clarke subdivisions, and advanced effects, and povray the lot at 24 frames per second, your cpu's can be pinned at 100% for literally hundreds of hours at a crack. My single 1.8 GHz processor can easily be pinned working on the same job for months on end (6 at least). Double the processor speed and you could look at 3 months. Now divide by 8 processors, 90 days turns into 11.25 days --pinned at 100%. Now I take the animation, and add 3 more scenes, and we are back up to 45 days of rendering with 8 cores twice as fast as what I am running now. There are literally a million computer applications that suck time hard. Over at Pixar, one frame from Finding Nemo took 4500 computers over 90 hours to render. Supercomputers with hundreds of thousands of processors (BlueGene/L, etc.) are usually capped to not run jobs that take more than two weeks to run. Short answer: they did not try very hard to 'max the processors'.
It's actually really easy to do if your memory system isn't meant to service 8 cores. And the article pretty much backs this up, every time the quad cores fail to shine it's blamed on the memory. But to me, the really interesting aspect of this is that they always blame FB-DIMM, which gains bandwidth by sacrificing latency. They even go so far as to suggest:
So, I think regular DDR2 @ 667 = 5.4 GB/s... divided amongst 8 cores is just 677 MB/s per core. It seems insane to think that would work (maybe it would, maybe my numbers are wrong also). If you want to attack latency but simply can't give up the bandwidth, wouldn't the SMP model work better-- just swap out the L2-miss stalled thread, and run the other full bore. Now you've reduced the problem to distributing your register bank among active threads. Well, I think that's how video cards do it, and memory latency is their enemy #1.
In any event, there you have it. The performance pendulum has left Ghz, is briefly swinging toward more cores, but appears headed now toward memory systems. Does anyone else think it's funny that L1 is still just 32kb? (oughta be enough for anybody).