Ars Technica Interviews 970 Designers

11223 writes "John "Hannibal" Stokes has interviewed Pete Sandon, the PowerPC 970's main designer, and David Edelsohn, a compiler writer from IBM, and clarified several points about the 970 regarding group formation, vector issue queues and performance, and more. The interview is a very interesting read for anyone who has been following his earlier articles on the processor that Apple calls the G5."

Comments

Altivec execution

[BWJones]

I was interested to find out find out they used the older Altivec instruction unit rather than the one from the G4e. Is there anyone that can comment of differences between the two Altivec units?

Windows based 970?

[Bill, Shooter of Bul]

I think if you read the end of the article where they are talking about the possibility of straight non apple boxes with 970 inside, you'll notice that they can not reply. Why not? It would be obvious to have linux based servers on top of the platform, so to have no responce to that makes me wonder if they aren't talking to soemone else about something nonobvious. What is the most non obvious step that would really get it in trouble with apple? Another deal with Microsoft. Heck the NT Kernel is portable and is currently being ported to itanium2 and amd64 why not ppc 970? I don't know how closely apple has tied panther to Their chipset, but if it isn't too tight this could mean cheap apple clones( they wouldn't ship with osx, but it could be installed). Now that would kill apple, and as a guy who had advance knowledge of it, I would simply say "No Comment" when asked about non apple based ppc 970 platforms.

Re:Windows based 970?

[pv2b]

Not legally.

You still need a licence to run Mac OS X, and I think it would be trivial for Apple to add a clause (if it's not already there) that would forbid installing the software on non-apple hardware.

There is also a port of Mac OS X for Intel processors being maintained in parallell, mainly because it CAN be done very easilly with minimal effort. Covering all bases...

Remember, Apple is a hardware company, and will protect their core business, which currently is and always has been hardware.

Re:Windows based 970?

[sergeantmudd]

Windows on 970 would kill Apple?? Hell, it would save Apple. How many times, on slashdot even, have you heard "I would buy OS X if the barrier to entry wasn't so high." The number one complaint against Apple isn't high price, it's high priced systems that aren't flexible. If I have an expensive PC with Linux and I end up hating Linux, I can put one of a bizillion different OSes on it. But I can only put a handful on the PPC box, and I can't run Windows. But if WindowsPPC becomes a reality, you will start hearing "Hey, I'll buy Apple and if OS X doesn't work out for me, I'll throw Windows on it"

Re:Windows based 970?

[Bill, Shooter of Bul]

Yeah, thats what I'm saying. What would stop people from doing it ( if its technically possible some people will do it regardless of leagality)? Here's a riddle for everyone to think about: If apple's a hardware company what unique piece of hardware does it actually make? The motherboard?
I think apple is really a software company that has managed to force people to buy hardware from it( at high margins) so their software can be run. As obvious as that statement is, I think many people forget it.

Re:Windows based 970?

[Johnny Mnemonic]

Now having "generic" white box 970's designed to run Linux is a different story.

Hmm...or IBM branded, consumer or corp grade desktops, with 970s, running Red Hat, Yellow Dog Linux, or IBMnix? I think the big advantages would be IBM QA and warranty on the hardware, and a linux optimized for the 970. Commodity parts, no M$ tax, the processor at cost since it's being sold by the fabber=cheap cheap and powerful. Say, $4-500 for a very competent office/workstation machine.

I can see these on a lot of desktops moving from Windows to Linux...I wonder if IBM knows any?

Re:Windows based 970?

[TheNetAvenger]

The NT kernel WAS portable. It has been said by MS kernel engineers that with Windows 2000, the changes made to the kernel made it very non-portable.

I'd like to see this quote, since the majority of the Win2k development cycle (including up to beta 2) supported the Alpha CPU.

So if Microsoft made NT (Win2k) non-portable did they do it in the Release Client a month or two before release just to screw with people or maybe in a mysterious service pack? Give me a break.

Win2k WAS designed for FULL Alpha support, and since the ALPHA development was cancelled LATE in the Beta, I seriously doubt they went back to the drawing board and rewrote the NT kernel after 10 years of portable design just to lock it into the x86 platform.

Besides, do you NOT get that the Itanium support in WindowsXP is not BASED upon x86, but a NATIVE NT Kernel running on the Itanium CPU, which has been available for almost two years now? Additionally the Windows 2003 Server support for the Itanium is a FULL 64 bit OS with a 64bit NATIVE kernel sitting on the Itanium processor. Geesh.

From the developers we worked with at Microsoft in the past year, even Windows 2003 server (which would include Win2k and XP) still have the basic HAL underpinning design and the lower NT kernel itself is STILL written in portable C.

According to the 'official' NT kernel development team, NT has been portable, stands by a portable design, and STILL is fully portable even TODAY.

So if you have some magical information that even the NT designers don't have, please share...

Improvements to GCC?

[Realistic_Dragon]

At one point in the interview it looks like IBM and Apple are working together on GCC improvements and donating the code back to the FSF.

This is a fairly big deal as people have pointed out before that GCC on PPC isn't as hot as it should be, but with that kind of muscle and money behind it it should go forwards by leaps and bounds.

With the new GCC improvements it looks like Linux on those new, remarkably cheap, P970 IBM boxes is going to be a real winner. And AFAIK Gentoo already runs on PPC fine - no one is going to be bitching about compile times with 4 1gig+ CPUs crunching away at it!

Re:Improvements to GCC?

[__past__]

"Donating" the code back to the FSF? They pretty much have to release it under a GPL-compatible license anyway, if they don't plan to keep their improvements inhouse.

Apple (or probably rather NeXT, but who cares nowadays) already tried once to improve GCC without releasing their changes, esp. the Objective C support. Turned out not to be such a good idea.

Re:Improvements to GCC?

[cactopus]

This box -- i.e. even future inexpensive ones are really designed to be Itanium and Xeon killers in addition to Sparc killers. They're basically going to be priced in the 3G-10G range for quite a lot of 64 bit computing muscle and the ability to run AIX. When you outgrow them you can move to POWER4+ on pSeries quite smoothly. A trusted and proven architecture.

These are basically this generation's Alphas... but with a better market positioning and without Digital/Compaq/HP at the helm. (we all know their pilot is dead at the wheel)

Re:Improvements to GCC?

[WatertonMan]

The issue isn't whether they'd release changes to the FSF. The issue is what to do with changes that the FSF doesn't want.

It may well be that the changes that IBM and Apple want to do to gcc are such that it would violate the basic model and methadology that gcc is following. At that point IBM can simply stick with Visual Age, accept a flawed compiler, or go with a fork. Apple's in a bit more of a pickle due to not owning Visual Age. There are rumors floating around about a port of Visual Age for OSX. However the problem of obj-C and so forth makes me think that this wouldn't help Apple as much as many think. Also it is in IBM's Linux plans to have a good gcc to make porting easier. So I think both want gcc to be as good as possible.

Even if they make a fork, I think everyone using PPC chips would use the Apple/IBM fork. So, outside of some likely contention with FSF, I don't think this would really cause that many problems. And the benefits definitely outweight the costs. (Assuming FSF doesn't get too pissed)

Re:Improvements to GCC?

[Big Jason]

As a Sun fanatic, I must admit that the POWER architecture is pretty nice. AIX OTOH is a steaming piece of garbage, and I'd rather be a Wintel admin than have to deal with it again. Hell, HP-UX 9.0 would be a step up.

I really wish Sun would ditch the ailing SPARC line and adopt the POWER/PPC line.

Re:Improvements to GCC?

[Duncan3]

Yes, it is great. Unfortunately in the next couple questions he goes on to say that the underlying engine GCC uses is completely unable to deal with the complexities of the 970 and other deep and wide architectures.

So either GCC needs a serious reworking on a fundamental level, or more likely since it already exists, they will just release a separate compiler that doesn't suck.

I just hope they release the proprietary compiler for OS X sometime before the G5 hits. The 2-3x performance hit of GCC is really starting to hurt Apple, and piss off all the developers. If IBM only releases the good compiler for their OSs and Linux, they are effectively telling people that OS X is not welcome on the 970.

Noone seriously considers using gcc on x86 now that the Intel compiler is free (root beer). And frankly that alone makes AMD chips totally unattractive for a computation farm (that and the nuclear plant you need to power the things).

IBM needs to step up and do the same.

Intel

[gilesjuk]

Wonder how long it will be until Intel headhunt these guys?

Wouldn't be the first time

Link to story

Re:IBM in the apple.slashdot.org section ?

[valkraider]

Why is Apple+IBM such a new thing?

How about October, 1991?

More info about the PowerPC alliance.

Apple never said that IBMs *technology* was no good. They said that IBM made boring corporate Personal Computers that didn't foster individuality and creativity amongst it's users. I would say that is still correct.

GCC

[eadint]

So whats the deal.
1) would it be possible to fork GCC so that it is completely optimized for the 970, if this is possible than GCC will be a good compiler.
2) from personal experience GCC sucks, its inefficient and wastes cycles, borland or Solarises compilers beat the crap out of GCC on their respective platforms. so when is apple going to have a highly optimized 970 based compiler that will make x86 whine in the corner after being raped, and made to look like the bad market whore that it is.
3) if a GCC970 can be separated, optimized and made into a special compiler, than well have something t talk about.