News for Monday, 25 August 2003

Thanks to all of those who pointed us to the release of Logic Platinium 6.2. This free upgrade brings mainly the 64bits optimizing for the PowerMac G5 processor, but also bug corrections, to enhance performance and stability when saving many audio tracks simultaneously.
in fact, not only is Logic Audio 6.2 optimized for the G5 but if you have this machine, you must have this version ! Hence its release at the same time as the availibility of the new PowerMacs.

As announced earlier, the sole 1.6 Ghz G5 started to hit stores. The Bad Duck saw it in the FNAC Champs-Elysées :

Hello, Today, visiting the Fnac in the Champs-Elysées, I unexpectedly came to find a G5 there. it was a 1.6 (with FX5200, well, bottom of the line model) with two small (so to speak) Apple flat screens linked on it. The storekeep was a real nice guy, we talked along for 10 mn, he let me see the inside and all, pretty clean, a really beautifully crafted machine :))) As for the noise, he told me (hard to make one's mind in the Fnac) it was just a bit noisier than an iMac. GGRRR, WHEN WILL THEY ARRIVE ! PS: I asked him when they would ship, he didn't know for sure, but didn't believe Apple's schedule would be right

Were Apple's planning right, the first next boxes should arrive at resellers' and wholesalers' who directly work with Apple on Tuesday.
[Upd]
After reading this news, Adrien went to the FNAC Champs-Elysées to try and see the beast. Unfortunately, it was already gone !
According to the storekeep (should be checked later on), it was the only machine in Europe ! And no more should arrive before the end of September.
So the G5 might still be Europe's most expected object for weeks...

According to informations we received the G5s are delayed because of a problem with the door lock that has appeared at the end of the assembly line. Many machines wouldn't pass the final test thus starting back again.

So as to get more info on the results realised by the G5, I managed to interview Kakace on IRC, kakace being the author of those parts of the client dnetc that contains optimizations for G4, G5 and Altivec. Let me remind you that the RC-72 challenge consists in gathering the most resource to break a 72 bits encryption key.
Why do you think the G5 is less performing than the G4 in your view ?

In bold terms, a G5 1.6 GHz keyrate levels the one of a G4 1 GHz (a 1.25 G4 reaches 13.3 MKeys/s per CPU). Besides that, I tried to write a specific core for the G5, but I have to admit it wans't taken into account (slower than core G4e). I need to do a new cruncher for the G5 (much more complicated). As my code generator actually doesn't enable me to do this, it will take a longer time.

What are the weaknesses of the current version ? why doesn't it work ?

The G5 AltiVec units have a slightly longer pipeline, which makes them very sensible to interacting instructions. Thus with a classic AltiVec cruncher, the G5 runs to a 60% level in regard of a G4 of a same frequency. To counteract those interactions, I'll have to write a much more complicated cruncher (which I call a 4-pipe/3 stages pipelined core).

What gains could be expected from this new cruncher ?

The G4 gives one AltiVec instruction per clock cycle, whereas the G5 is closer from an instruction every 2 cycles. This new cruncher's aim is to eliminate the dependence links between instructions to reach full-speed. If I do my job as I wish, the G5 speed should overcome a same frequency G4.

How long did you need to elaborate the current optimization for G5 ?

Almost no time :) I wrote a tool to generate ASM code for G4s crunchers. I just modified a few tricks to try to fill the G5 pipeline, but this stuff didn't work as expected :-/

It seems to be a case of failure in optimization :-)

Yes, a failed adaptation. My goal was to improve perfs, as much as possible, without writing a new cruncher. An almost impossible task when you don't have a G5 actually :-) I did this blindly, after the 970 specs. That is to say, the current 970 cruncher is a modified G4e one, thus a non-optimum one

How will you write this new cruncher and how long will it take ?

As soon as the G5s will be delivered, I'll try to write a new, specific cruncher. I'll just have to give s.o. owning a G5 a test purpose client to launch a bench. From that point, I'll analyze those results with regard to the code and theoretical specs. I dunno how long this will take... everything has to be done again, this may require more than a month to complete.

Last question, what would a 64 bits optimization of your code entail, using this G5 64 bits nature ?

Nothing at all (on the contrary). The algorithm is made for 32 bits. In 64 bits mode, the resul of each operation has to be reduced so it is modulo 2^32, whihc means loads of time lost.

Which means ?

As the challenge consists in decrypting a message using a formula based on exclusively 32 bits calculations, it would take RSA to initiate a new challenge with a cryptage based on modulo 2^64 calculations.

Thx Didier for those precisions, we hope you'll son give us the cruncher that will enable Apple to justify their motto, "world's fastest personal computer".

In my view, a big G4 already does pulversih any thing. Just watch the stats on distributer.net to get aware of thta. So the G5 might only improve that result (even thru the sole frequency gain).

Those stats are available on distributed.net web site. You may also apply to RC5-72 challenge with team Macbidouille.

Always getting better ! Microsoft points a huge security leak in its system, this time because of Internet Explorer - but this time one may be concerned even without using IE. It is so much part of the system that its only presence will be a risk !
Send your PC-ist friends here