The 68040 was faster than the horribly designed 603 based PowerMacs, not all powerpc chips. As I said above, the 603 was really slow even for the time, it was a budget low power chip. Then on top of that Apple crippled them further by putting them in really shitty machines that slowed them down even more.

The 68040 was usably faster than the early 601 machines as well, but that was mainly because of the code emulation needed. With native code a good 601 machine was faster than a good 68040 machine.

For 604 and above, the powerPC was faster.

There were, but they were extremely rare. Apple was moving to powerPC so the upgrade chips did too.

Because:
a) The 68K was at the end of the line. Apple was too small at the time to keep Motorola interested in continuing to develop it.

b) RISC was "the future". The Alpha was kicking ass and taking names back then, it didn't matter that the rest of the RISC camp were no faster than the good CISC chips.

c) PowerPC offered the closest "drop in" replacement. It could emulate the 68K reasonably well and in some cases could use most of the same hardware, at least poorly. Apple saved craploads on R&D by taking the easiest path they had.

I'm kinda sad and disappointed Apple didn't opt for Dec Alpha

That probably would not have helped much, if at all.

The high end machines would have been insanely fast, but also even more insanely priced. The $5K power macs of the time would have been downright cheap in comparison.

The low end machines would have still been low end. Just like there are Celerons, there were low end Alphas. Not as bad as the 603 was, but Apple would have likely used lots of them. The low end Alphas got their asses handed to them by Intel chips just like the low end PPC chips Apple did use.

Even if Apple had used the Alpha they didn't have the volume at the time. The Alpha might have survived a bit longer, but in the end the result would have been the same: Apple running on Intel.

so um basically based on this history, how do you think arm vs intel will turn out? RISC may not be able to outperform intel on the high end but on the mobile end, the decode logic penalty creates a power drain.

I wouldn't bet against Intel, certainly not if they're on their A game. Their design, engineering and process is second to none. They may well screw up for other reasons of course, eg. marketing or playing silly buggers trying to get lockins.

I also wouldn't bet against ARM no longer being an independent company in ten years time - not because I think they're going to fail and get gobbled up, but because they may become strategically vital to someone with a lot of money, and get gobbled up. Apple, Samsung and Intel spring to mind. I know there's been talk of a poison-pill clause in their company statutes, no idea how that might play out. Nor competition regulation issues either.

The other thing is where are you drawing the battle-lines? Mobile phones? Tablets? Low-powered servers (where AMD appear to be going with ARM)? Laptops? Desktops? Each of these areas may well play out differently, with Intel clearly being at less of a power disadvantage in the latter four than in mobile phones.

By they way there seems to be a bit of a difference of opinion quite how significant the x86 decode penalty really is. In a world where ARM is going multi-core, out-of-order and 64-bit the extra bit of decode logic may not be such a big handicap outside the very lowest power devices.

I know I'll likely get flamed for it, but here goes:

The PowerPC, as used by Apple, was almost never actually faster than contemporary x86 chips. It was all Apple's marketing BS.

Except for a very brief period where the Power PC 604 was king before the Pentium Pro came out, the x86 world had the faster CPUs. Sure there were some corner cases when the Power PC could beat out x86 such as when code could take advantage of Altivec, but as general purpose CPUs the x86s won in almost all cases.

Apple was able to weasel around the slower CPUs somewhat by not running comparable OSes. MacOS before X was a pathetic excuse for an OS, even when it was current. It lacked many basic features that contemporary OSes had for years before. As such the OS was very simple and didn't need much hardware to feel fast. It would be like comparing Win 3.1 on a P1 to Win2K on a P2. The Win 3.1 box will probably "feel" faster, but that's because it's so simple compared to Win2K. MacOS was even simpler than Win 3.1. Even once OSX was released, it was so different than even other Unixes kernel wise that a direct comparison was impossible. Performance was "good enough" that they could get away with it.

If you paid attention during the G4 & G5 days you'll notice Apple only compared their machines to Intel ones. Not to AMD chips which were even faster at the time. They couldn't get away with stretching things that far.

If you want to prove this to yourself, just find a PC and Mac from the same era and install Linux or NetBSD on them. Then benchmark. Or just run some apps. Then realise how much BS Apple was spouting about the PowerPC being faster. I've done this with several generations of x86 and PowerPC CPUs. The x86 chips won handily except for the 604 vs P1 case above.

Another proof of Apple's BS marketing is the x86 Developer Preview machine. This was a bog standard P4 machine Apple sold to developers before the PowerPC->x86 transition. It's the same P4 chip that Apple claimed was so much slower than the G5. Once running the same OS the preview machine showed how wrong Apple was. Developers were raving about how fast it was compared to the G5.

Why was the x86 faster? First off, just because a CPU is RISC does not make it fast. ARM chips are RISC, but only recently have they been able to start giving even atoms a run for their money performance wise. Sun's Sparc based chips were put into awesome machines with multiple CPUs and I/O that put any PC of the time to shame, but on CPU vs CPU level they were often slower as well.

Second, CISC has some advantages. Code is usually denser because it can do more with fewer instructions. In the days when caches were small this was a real advantage. Even now it helps deal with the fact that RAM is still slow compared to the CPU.

Third, Intel and AMD have added pretty much every "RISC" feature to their CPUs even when the RISC crowd said it was impossible. Out-of-order execution, branch prediction, register renaming and speculative execution? The P Pro had all of those. Superscallar? P1 had that. Pipelining? Even the 486 had that. Many of the contemporary PowerPC CPUs didn't have all those features. The 603, 603e and G3 were used in many Macs, but they lacked many of the advanced "RISC" features that their CISC x86 competitors did have.

Basically every x86 CPU from the P Pro on has had almost all the nice "RISC" features as well as having the CISC instruction compactness. The best of both worlds makes for fast CPUs.

Now, there are some awesome RISC designs like the Alpha and the POWER line. They both (and still do in the case of POWER) ran faster than contemporary x86 chips. But they did it with extreme clock speeds and extreme levels of heat.

what about floating point, then and now?

i recall that RISC FPU performance was more than double Intel. Now I know x87 is deprecated and SSE/AVX has improved.

More than double is an exaggeration, if not an outright lie. Maybe close to true when the 486 was the best Intel had, but after the P1 and especially after the PPro came out it was a lie.

Again, not every RISC CPU was the Alpha. The PowerPC and SPARC chips were at best equals to contemporary Intels, but often slower. ARMs even to this day often lack FPUs entirely.

well wikipedia states

http://en.wikipedia.org/wiki/Pentium_Pro

Compared to RISC microprocessors, the Pentium Pro, when introduced, slightly outperformed the fastest RISC microprocessors on integer performance when running the SPECint95 benchmark,[6] but floating-point performance was significantly lower, half of some RISC microprocessors.[6] The Pentium Pro's integer performance lead disappeared rapidly, first overtaken by the MIPS Technologies R10000 in January 1996, and then by Digital Equipment Corporation's EV56 variant of the Alpha 21164.[7]

6-^ a b Slater, Michael (13 November 1995). "Intel Boosts Pentium Pro to 200 MHz". Microprocessor Report.


I do wonder how specFP of PowerPC G4 and G5 compare to the Intel equivalents.

Emphasis mine. Sure the Alpha and the MIPS at the time beat the PPro. That doesn't mean the PPro didn't beat the PowerPC.

Some SpecFP 95 numbers of CPUs available at the time:
PPro 150 MHz: 5.42
PowerPC 604 133 MHz: 3.76 Scaled to 150 MHz: 4.42

And that was the "fast" 604, not the much more common 603.

As a point of comparison:

Ultrasparc 1 143 MHz: 7.90 Scaled to 150 MHz: 8.28

The Ultrasparc 1 was faster at FP, but only about 1.5X, not "more than double".

Both the PowerPC and Ultrasparc 1 got beaten at specINT by the PPro, even if scaled to the same MHz.

The PPro actually beat the MIPS on it's release date in FP, but to be fair much better MIPS came out shortly after.

R5000 150 MHz: 4.78



I could not find Spec numbers for the G4, but some SpecFP2000 numbers for a G5 from late 2004: PPC 970 2.2 GHz 2 cores: 1241 Core scaling wouldn't be perfect, but that's about 621 for a single core machine.

It looks good against a P4 of the same speed, except that the P4 is single core: P4 2.2 GHz: 777

But it wasn't competing against the 2.2 GHz P4 in late 2004, it was competing against this (has hyperthreading, but single core): P4 3.4 GHz: 1561

And this single core CPU: Athlon 64 FX-55 2.6 GHz: 1878

I'd say "ass handed to it" is appropriate for the G5 vs x86 of the time

ok. wow. Is integreating the GPU ala AMD or AVX 3.2 ala Skymount further improve x86 specFP?

What happened to the promise of EPIC/Itanium? I'm well aware that compilers are hard to write for and the very first Itanium underperformed.

But Itanium was supposed to be a clean-room, no-legacy forward-looking CPU the only 64-bit for Intel, with Intel retiring x86 at 32 bit.

EPIC based on VLIW was supposed to be a leap beyond Risc/Cisc

The AMD way no, the AVX way yes. Despite the integration the GPU is still a separate "processor" in AMD chips. It does not run x86 code, even FP code. It may count towards the specFP of the system but not the specFP of the processor. AVX is part of the main CPU.


You said it yourself. The compilers couldn't optimize the code well enough. A compiler cannot optimize for every possible outcome. A modern CPU looking at the code in real-time can do a better job.