[Nerozwei]

Check out this link to see a comparison of PPC processors.

Let's see three competitors:

Power 4 (1x1.3 GHz *), score ~8M - 6.15 pts/GHz
970 (2x2 GHz *), score ~17M - 4.25 pts/GHz
G4 (2x1.25 GHz *), score ~14M - 5.6 pts/GHz

* I'm assuming the G4s and G5s are dual and the Power 4 is single

If I get it right bus speeds and such don't matter that much in this type of computing? The G4 scores that high because of Altivec, right? But why is the pts/GHz ratio of the G5 is that low? Longer pipeline, smaller cache?

The Power 4 is so strong because of it's cache?

Could someone wiser in CPU tech explain this to me? Maybe the results are just bullsh*t

EDIT:

Hmmh... I wonder if the Power 4 is even available in a single proc config....

[reader50]

The Power4 is a dual-core chip. Two CPUs per die, with insane interconnect bandwidth between them and 1.5MB of shared L2 cache. It doesn't have AltiVec.

[geekwagon]

Originally posted by Nerozwei:
Hmmh... I wonder if the Power 4 is even available in a single proc config....

It is. I can't remember about the 1.3GHz in particular, but I know they have a single-core version that they use in some of the lower end pServers (like the p615) and in workstations. Currently you can buy a 1.25 in a single core version...

The other thing it can be is that all modern Power4 equipped machines are capable of dividing their CPUs and memory up into LPARs (virtual machines, basically), so it would not be surprising to see an LPAR booted up with only one CPU in it.

I am not sure why the Power4 would be so much faster per MHz than the G5 at distributed.net, though. I have a 2-way p615 and a 8-way p650, each with 1.45GHz Power4+'s (fastest currently available) arriving next week. Maybe I should load up distributed.net and see how fast they are

[Nerozwei]

Originally posted by geekwagon:
I am not sure why the Power4 would be so much faster per MHz than the G5 at distributed.net, though. I have a 2-way p615 and a 8-way p650, each with 1.45GHz Power4+'s (fastest currently available) arriving next week. Maybe I should load up distributed.net and see how fast they are

Hmmh.. I assume they're not just your personal hobby boxes What will the Power4+'s be used for? And if possible, try out the dnet client

[djohnson]

You must remember that the G5 client is not optimized. No telling when they will get around to it either...

[djohnson]

One other thing, the speeds for the G5 are per processor.

[Lateralus]

Speaking of Distributed:

I may or may not be snagging up one of those shiny Dual G4 processor uprades for my machine soon, and I was wondering if it is possible to run the SETI client on one processor and Distributed on the other? I know you can run two instances of SETI, but I have never heard anybody mention anything about two different clients being run simultaneously.

Thanks.

[Scotttheking]

Originally posted by Nerozwei:
Check out this link to see a comparison of PPC processors.

Let's see three competitors:

Power 4 (1x1.3 GHz *), score ~8M - 6.15 pts/GHz
970 (2x2 GHz *), score ~17M - 4.25 pts/GHz
G4 (2x1.25 GHz *), score ~14M - 5.6 pts/GHz

* I'm assuming the G4s and G5s are dual and the Power 4 is single

If I get it right bus speeds and such don't matter that much in this type of computing? The G4 scores that high because of Altivec, right? But why is the pts/GHz ratio of the G5 is that low? Longer pipeline, smaller cache?

The Power 4 is so strong because of it's cache?

Could someone wiser in CPU tech explain this to me? Maybe the results are just bullsh*t

EDIT:

Hmmh... I wonder if the Power 4 is even available in a single proc config....

The numbers are all for single procs. As for cache and bus speed, none of that matter to the dnet code. It loads itself into L1 cache and crunches away. Since it doesn't actually do anything, it has no data to cache or transfer from memory.

[geekwagon]

Originally posted by Nerozwei:
Hmmh.. I assume they're not just your personal hobby boxes What will the Power4+'s be used for? And if possible, try out the dnet client

No, they are for work, I don't have the $150k to throw around that those things cost.. The larger box is going to be a data-warehouse server running Oracle, the smaller one is just a box that I can use to compile software that I need for AIX and other management tasks.

These boxes aren't going into production until second quarter.. Maybe I will load the dnet client on them to "test performance and burn in the CPUs"

[Nerozwei]

These boxes aren't going into production until second quarter.. Maybe I will load the dnet client on them to "test performance and burn in the CPUs" [/B]

Ahh.. the classic burn in excuse You should really try distributed. Just imagine your geekfactor rising...

[Nerozwei]

Originally posted by djohnson:
One other thing, the speeds for the G5 are per processor.

Mmmmm... I wonder how the Virginia Tech would do in Distributed net? 1100 * 2 * 19M. Might just kick my G3s buttah

[i_wolf]

i could be completely misinterpreting the results. I am assuming the 970 G5 results are single cpu results. I never knew about that bench. But if anything is the G5 not the fastest on that link???
It would be very interesting to see how both the Power 4 and G5 would bench if they compiled with an optimal compiler.
Do these bench's take into account altivec???

[Nerozwei]

Originally posted by i_wolf:
Do these bench's take into account altivec???

I believe it does. I think there were credits shown about Altivec implementationm etc. Around four years ago I remember a slow G4 (as in MHz) kicking fast PCs asses.