The Sony E3 megaton is ... Apple?

[Entropy]

Sharp eyes chachi.
I've been turning this over in my head quite a bit before. While there is little doubt that Merom/Conroe will offer good performance both per Watt and in absolute terms, that "integer performance" in the graph is critical.

How important will integer performance be in the future?
Amdahls law applies to all aspects of architecture, i.e. if your architecture has a weak spot, that property will typically prove limiting for your code be it scalar performance, bandwidth, branch mispredict penalties, whatever. So you can't neglect integer performance too much. But as far as I can see, the stuff where an end user actually percieve that they sit around waiting for the CPU is mostly FP heavy. Now, even FP heavy code includes a fair bit of integer operations, but still.

The new console cores from IBM make their performance trade-offs in a non-traditional way compared to x86. They trade integer single thread performance for good multi-thread and FP performance, and to my mind that seems pretty reasonable going forward.

I'm still a bit mystified as to why Apple didn't go with something like the XBox CPU, perhaps with small variations on that theme (dual core/larger caches?) and revised I/O. Sure, they would loose a bit in terms of single thread integer performance, but would that be a big deal compared to the gains? Perhaps the production volumes would be too small, but isn't IBM touting its willingness to spin dedicated silicon for its PPC customers? Hmm.

 

[Gubbi]

But as far as I can see, the stuff where an end user actually percieve that they sit around waiting for the CPU is mostly FP heavy. Now, even FP heavy code includes a fair bit of integer operations, but still.

What are these workloads ? HD-video and games? Both of these are well suited for GPU accelleration.

It's not like future x86 will be lame in SP FP. Micro-op fusing and full width SSE units will allow 1 SIMD FMADD per cycle, quadrupling performance compared to today... per core.

In two years time, PC CPUs will be comparable to XeCPU in FP and alot better at executing sequential spaghetti.

The new console cores from IBM seem to make their performance trade-offs in a non-traditional way compared to x86. They trade integer single thread performance for good multi-thread and FP performance, and to my mind that seems pretty reasonable going forward.

Hyper-pipelined super high clocked CPUs is not the way to achieve lower power.

Console CPUs:
1. The XeCPU supports 6 threads, which is nice.
2. CELL, two threads in the common multi-processing OS sense, and a bunch of disastrously difficult-to-program co-processors.

The XeCPU might have made sense in a Power Book, if the power consumption were low enough and integer performance up to par, which it isn't. But XeCPU is the (IP) property of MS and hell will freeze over before they'd grant a license to Apple.

CELL is just too special purpose to go in any desktop or laptop. A big expensive piece of silicon of which 8/9 will sit idle for most of the time.

Cheers
Gubbi

 

[Entropy]

It is significant to note that OS-X has been x86 native since its inception. This is indicative of some damn good management IMO, and it makes me inclined to believe that Apples reasons, whatever they are, are pretty good. They have been able to pull this one for some time as well as postpone it indefinitely if they so chose.

I still don't see the need for higher integer performance than a rehash of the Xenon CPU would offer though. I happen to type this on an 800MHz iMac, and the applications Gubbi pointed out, HD-video and games, are the only instances where I ever feel that this machine might need more horsepower. And the 800 MHz G4 is dog-slow compared to everything we're discussing here.

Scientific/technical computing will always have particular needs, but will also be able to fund the hardware needed. Those applications don't have any bearing on the direction of desktop computing, really. So media is what requires home computing horsepower today, pretty much exclusively. To my mind, that implies that using an architecture tailored to suit those applications would be optimal. Performance on legacy and office/adminstrative apps just isn't the issue anymore.
I quite understand this move on the part of Apple, I just regret that something architecturally more radical wasn't selected.

 

[Fox5]

I'd imagine Apple starting eyeing x86 cpus when both the Athlon and the P3, and later the P4, soared beyond the G4 in performance. That the G5 barely manages to remain competitive and Centrino removes all hardware reasons to get an iBook probably helped reason the move to x86.