Welcome to the MacNN Forums.

yoyoman · Jul 19, 2003, 05:29 PM

barefeats says 7/19/03 -- One of my loyal readers ran AltiVec Fractal on G5/2GHz MP at the NY MacWorld. With the count set to 65K, he got 10.5 Gigaflops. On the G4/1.42GHz MP in our lab I get 10.8 Gigaflops. Don't jump to conclusions, but this might explain why the G4/1.42 MP disappeared from the Apple Online Store the day the G5 was announced. ;-)

How odd is that.

Ganesha · Jul 19, 2003, 05:32 PM

Originally posted by yoyoman:
barefeats says 7/19/03 -- One of my loyal readers ran AltiVec Fractal on G5/2GHz MP at the NY MacWorld. With the count set to 65K, he got 10.5 Gigaflops. On the G4/1.42GHz MP in our lab I get 10.8 Gigaflops. Don't jump to conclusions, but this might explain why the G4/1.42 MP disappeared from the Apple Online Store the day the G5 was announced. ;-)

How odd is that.

Not only that! The G5 is slower then the IBM XT! The CPU and BUS run at the same speed on the XT! This may explain why the XT is disappeared from the retail market!

teh tenk · Jul 19, 2003, 05:55 PM

Also notice how gigaflops has disappeared from the Apple vocabulary.

The new buzzword is "bandwidth", man. Get with it.

alien · Jul 19, 2003, 06:32 PM

There are some performance differences with both AltiVec and cache instructions between the 745x and 970 processors. Software needs to be re-tuned for the 970 to get optimal performance.

Some cache/prefetch instructions that can help increase performance on the G4, can be big no-no's on the G5.

The tester at MWNY probably used the G4-optimized version of the AltiVec Fractal demo. Let's wait for a newer version...

Eug Wanker · Jul 19, 2003, 06:57 PM

There is a thread over at Ars discussing Xbench Altivec results. The basic consensus (after analysing the code) is that the bench is completely inappropriate to assess comparative Altivec performance.

I wonder how Altivec Fractal would hold up under the same scrutiny...

Scotttheking · Jul 19, 2003, 08:54 PM

Originally posted by Eug Wanker:
I wonder how Altivec Fractal would hold up under the same scrutiny...

Run Shark on it and see.

Eug Wanker · Jul 19, 2003, 09:07 PM

Originally posted by Scotttheking:
Run Shark on it and see.

I'll leave that up to you. I'm not a programmer, just an end user.

Thinkdiff · Jul 20, 2003, 12:05 AM

A G5 will probably not run AltiVec code as well as the G4. The 970 was made to run 64-bit code optimized for it, not AltiVec instructions. That was added later. The G4 was made to run AltiVec, therefore making it faster. You will have to use a 64-bit coded app optimized for the G5 to see any real gain.

DaedalusDX · Jul 20, 2003, 01:19 PM

Originally posted by Thinkdiff:
A G5 will probably not run AltiVec code as well as the G4. The 970 was made to run 64-bit code optimized for it, not AltiVec instructions. That was added later. The G4 was made to run AltiVec, therefore making it faster. You will have to use a 64-bit coded app optimized for the G5 to see any real gain.

Uhhh... There is some validity with what you are saing, but I think you are a bit misinformed.

The G5's altivec unit was probably added to the design inherited from the Power4, and there are aspects about it that are inferior to the G4's altivec.

For a more in depth analysis, read here: http://arstechnica.com/cpu/03q1/ppc9....html#970vsg4e

However, there are MUCH more factors than simply the G4 was designed with altivec more in mind. The G4 is also crippled with a 167 Mhz front side bus. Lots of times, while doing heavy altivec instructions especially, that bus is saturated, and the processor is not running at nearly its potential.

The G5, on the other hand, has a 1Ghz front side bus with LOTS of bandwidth.

In addition, Altivec is only one aspect of performance. Only in heavily altivec optimized applications will there be a difference. For every other instruction used in any other application, the G5 will perform better than the G4 simply because its a more advanced architecture with lots more bandwidth.

Being a 64 bit processor doesn't mean it will not run the same 32 bit PowerPC code any slower. In fact, that is were most of the performance gains will be seen... regular 32 bit PowerPC code.

64 bit will be nice, but that type of stuff will likely be used only for specific high end apps. Even after the G5 has been around for a while, most code will still be 32 bit. This is definitely not a bad thing though. It will run the 32 bit code faster.