About PPC and x86 performance at a glance

[patroclus02]

These days I've been trying to get a overall idea of the performance of PPC architecture against x86, now that Mac uses Intel CPUs. I'm so familiar with many processors at a quite deep level, but never minded so much about PPC world.

Of course, it is so hard to get exact results, and depends a lot on the applications, and the rest of the system configuration, and I'm only searching for a very general view.

We could say G3 is a little bit slower than a Pentium III.
G4 becomes a little faster than a Pentium III and at Athlon level at same clock freq.
G5 is just a little bit faster than G4 at same clock, but reaches higher frequencies.
So, it could be at an Athlon XP level (always at same MHz)
But Core Duo and Athlon 64 are a bit faster than G5, and Core 2 Duo quite much faster.

Could this be considered correct?

 

[Saem]

Ok, this doesn't make sense, are you comparing dollar wise? If so, everything x86 is far faster as you can get much faster CPUs at the same price point, except for rare edge cases (a subset of which were actual applications), and Apple FUD were about the only time PPC was really any faster.

If you're using clockrate then PPC systems could maintain a higher IPC, too bad instructions aren't all that comparable.

An easy way would be to compare Spec benchmarks, but the problem there is that's more whole system, and the double precision FP stuff really made the G3, and G4 look bad, IIRC.

 

[patroclus02]

Of course, I understand. I mean raw power.
I looked lots of benchmarks (everyone I could find). I just wanted to get a very aprox idea of where to place them all

 

[Fox5]

Of course, I understand. I mean raw power.
I looked lots of benchmarks (everyone I could find). I just wanted to get a very aprox idea of where to place them all

Raw power is a suprisingly undescriptive term.

Considering that for the most part PPC and x86 never ran the same software, and definetely not with equal levels of optimization, it's really difficult to place these.

You're probably approximately right with your ideas, but there's no definitive answer as it will vary from application to application, you're not using exact terms anyway, and ultimately it doesn't really matter.

 

[ShaidarHaran]

If you want to define your query a bit better by stating "peak GFLOPs or Integer ops/s" then it becomes easier to answer. IIRC the fastest x86 MPU has a peak of 96GFLOPs (QX 9770). If we consider Cell to be the fastest PPC MPU then its 8 SPEs @ 3.2GHz are capable of delivering 204.8GFLOPs peak.

 

[Fox5]

If you want to define your query a bit better by stating "peak GFLOPs or Integer ops/s" then it becomes easier to answer. IIRC the fastest x86 MPU has a peak of 96GFLOPs (QX 9770). If we consider Cell to be the fastest PPC MPU then its 8 SPEs @ 3.2GHz are capable of delivering 204.8GFLOPs peak.

From his original post, I think he was only referring to PPC cpus used by Apple, thus IBM's power series and cell are out of it.

 

[Snyder]

If you want to define your query a bit better by stating "peak GFLOPs or Integer ops/s" then it becomes easier to answer. IIRC the fastest x86 MPU has a peak of 96GFLOPs (QX 9770). If we consider Cell to be the fastest PPC MPU then its 8 SPEs @ 3.2GHz are capable of delivering 204.8GFLOPs peak.

But the SPEs don't have much in common with PPC ISA, do they?

 

[ShaidarHaran]

But the SPEs don't have much in common with PPC ISA, do they?

Nope, It's hard to find peak rates for any purely-PPC MPU that's not in a server though. so Cell was the closest I could get that seemed relevant.

 

[chavvdarrr]

If you want to define your query a bit better by stating "peak GFLOPs or Integer ops/s" then it becomes easier to answer. IIRC the fastest x86 MPU has a peak of 96GFLOPs (QX 9770). If we consider Cell to be the fastest PPC MPU then its 8 SPEs @ 3.2GHz are capable of delivering 204.8GFLOPs peak.

that will be full precision vs single , Cell tops at 20-30 GF fp64

 

[ShaidarHaran]

that will be full precision vs single , Cell tops at 20-30 GF fp64

Then take HPC Cell which tops out at around 100 DP GFLOPs (half SP rate for DP).

 

[patroclus02]

I know my question was far too ambiguous (yes, I meant Mac PPC).
I suppose I was interested in apparent power the user perceives.
There're similar applications for both, Mac OS X and Linux/Windows, probably with a different level of optimization (for example, Photoshop performance in PPC is something almost unbelievable), but if you get a Mac or a PC for general purpose/internet or all-task computer, you always get an apparent performance. Is like when I go my Duron 700 MHz/128 Mb, to replace my K6-2 300 MHz/64Mb. It seemed to fly

 

[hoho]

It was VMX that made PPC so powerful. Sure, Core2 and K10 come close (if not exceed) in SIMD performance but they don't yet have the flexibility thus there are some cases where PPC can still vastly outperform them (see previous comment about Photoshop )

 

[Thorburn]

(for example, Photoshop performance in PPC is something almost unbelievable)

From what I recall Photoshop CS2 on a Windows XP/x86 system would generally be faster than a OS X/PPC machine...

OS X/x86 vs. OS X/PPC is a different story as CS2 wasn't a universal binary and so ran via Rosetta.