Even though Amorph and Wizard69 did a fair job in getting something remotely credible out of this rumor, I still find this rumor to have way too many unanswered questions about it.
How will they get Altivec to perform acceptably?
Like it's been proposed now, we wind up with 4 Altivec-units. This would be pretty equivalent to a very low-clocked, but flexible 512-bit Altivec.
Even 256-bit Altivec has been doomed into the land of the law of diminishing returns, because it will be very hard to find enough parallelism in the code to exploit this (and it would make the unit itself a transistor monster, but that's not the concern here).
Yes, you could probably get acceptable performance out of these four units, if you're skilled enough as a programmer to manage to split the code onto all the units, but I believe that's pretty hard, and more importantly: It would require a rewrite of all code.
On regular code, optimized for the 970, and old (legacy) AV-code, it would perform at ~30% compared to the 970, because of the low clock-frequency, which is pretty low in my book.
This could be possibly be solved with the technology that would have to be the key to the biggest problem of this design entirely: To make it possible for all the cores to execute on the same thread, i.e. eliminate the need for multithreading. (And don't ask me how to do that, if it's even possible)
Without this, this pseudo-G5 Powerbook would be a terrible machine for all things involving less than 4 threads. A typical example of this is benchmarks. Benchmarks are almost always performed with the benchmark itself as the only running CPU-intensive task. And the typical benchmark (especially the cross-platform ones) is not multithreaded.
So unless some breakthrough is achieved relative to multi/single-threading, this machine would probably stand out as very slow in every aspect of benchmarking, except where they tested multitasking extensively.
this pseudo-G5 Powerbook would be a terrible machine for all things involving less than 4 threads. A typical example of this is benchmarks.
Apple is the only personal computer company that can change the way computers work. Changing the Power Mac line to dual-processors kicked in SMP on the Mac platform. Cell-style quad-core PowerBooks would further increase the number of threads, and more importantly the efficiency of threads (lot's of programs unfortunatly have token multithreading, I've seen).
Also, I don't give a rat's arse about benchmarks. Let Apple thread their apps better and only give benchmarks for them if necessary.
Apple is the only personal computer company that can change the way computers work.....Also, I don't give a rat's arse about benchmarks. Let Apple thread their apps better and only give benchmarks for them if necessary.
I don't know, all this story is confusing. I believe Zapchud has valid points, though I would like to see a powerbook with the processor discussed here. Certainly, Apple will thread further their applications to take advantage of the new architecture, if this happens. And i would prefer to use such a machine over others, since the architecture seems efficient and potentially more reliable than what is available today in the notebook market. Use of cell-like processors by Apple would mark a new era in personal computing.
however as with any revolutionary advancement in computer design, there will be a period in which migration to new programming model is require
apple and IBM have been porting os x to the new programming model.
however
even with single thread, performance will be acceptable for mobile computer. the computer will feel responsive, and processor intensive applications will probably be quickly ported.
Don't know, but wouldn't adding a floating point unit throught the existing "Auxiliary Processor Interface" on a 440 core, increase some kind of latencies or inefficiencies?
Nr9
Do you have actual inside information or is this all just speculation?
And you're sure this is for a Powerbook? Seems multithreading and multitasking is advancing rapidly, but for this type of system to compete with the brute force of a 9 to 20 stage piped G5 stills seems far fetched to me.
The key thing here is responsiveness, and this is where Nr9 really has a point. Anyone who's used a duallie remarks on how silk-smooth the interface is. Go up to four processor cores, and you have a machine that can multitask up a storm and still be instantly responsive. Really, the programming model only becomes a big issue if you decide to run a giant single-threaded beast like the average ported video game.
Engineering is always about tradeoffs, but if you're going to have pervasive tasking and threading - as OS X does - then many hands make light work of it.
The only remaining champion of the traditional personal computer architecture is Intel.........This sort of 440-based board is one of the mammals scurrying around under the feet of the mighty dinosaurs........
Not that my opinion is worth anything, but I would like to state that this is the best, most thought provoking post I'ver ever read on AI. And it manages to include: a dinosaur analogy and the absolutely grammatically correct spelling of 'de-emphasize' (go and look!)
Regarding FPU through the auxillary processor unit interface, this is already known to be working and available from IBM. Also, their latest PPC newsletter has lots of ppc 44x articles, especially the new product focus. Specs for the 440GP are eerily familiar to what Nr9 has described.
According to roadmaps I've seen Mojave will be here in 2004, and Orion in 2005.
Arent Orion something with stars (And a Metallica song) and not a desert?
Quote:
Originally posted by Nr9
even with single thread, performance will be acceptable for mobile computer. the computer will feel responsive, and processor intensive applications will probably be quickly ported.
Arent Orion something with stars (And a Metallica song) and not a desert?
It's a contellation yes. Orion is not a codename for the 7xx-series witch seem to have a desert theme to its codenames (Sahara, Gobi and Mojave), it's a codename in the 4xx-series witch seem to have heaven-related codenames (Aurora and Orion).
And.. Amorph's and Barto's diagrams were very beautiful! I'm not worthy
Comments
Originally posted by murk
If this is true and it will be called a G5, does this also mean that the 970/980 will also move to multi-core?
Doubt it, at least not for the 970. As it is, there isnt a need to move in that direction.
Right now the best thing that could happen to the 970 is a die shrink and a clockspeed of 3Ghz.
i love threads like this, learn so much.
Originally posted by wizard69
Frankly this looks like the technology that should be going into an IBook.
Do you want an iBook that costs more than a PowerBook?
Thought not. This is a cool direction, but as with any new direction, expensive initially.
Barto
Originally posted by Amorph
My 3am attempt at a logical view of the layout...
Has the new dual G4 rumor some relation with what it is discussed here? What do you think?
Originally posted by Nr9
sory i had typo. it is one MCM with two chips that each contain two processor package together
More like this.
Barto
How will they get Altivec to perform acceptably?
Like it's been proposed now, we wind up with 4 Altivec-units. This would be pretty equivalent to a very low-clocked, but flexible 512-bit Altivec.
Even 256-bit Altivec has been doomed into the land of the law of diminishing returns, because it will be very hard to find enough parallelism in the code to exploit this (and it would make the unit itself a transistor monster, but that's not the concern here).
Yes, you could probably get acceptable performance out of these four units, if you're skilled enough as a programmer to manage to split the code onto all the units, but I believe that's pretty hard, and more importantly: It would require a rewrite of all code.
On regular code, optimized for the 970, and old (legacy) AV-code, it would perform at ~30% compared to the 970, because of the low clock-frequency, which is pretty low in my book.
This could be possibly be solved with the technology that would have to be the key to the biggest problem of this design entirely: To make it possible for all the cores to execute on the same thread, i.e. eliminate the need for multithreading. (And don't ask me how to do that, if it's even possible)
Without this, this pseudo-G5 Powerbook would be a terrible machine for all things involving less than 4 threads. A typical example of this is benchmarks. Benchmarks are almost always performed with the benchmark itself as the only running CPU-intensive task. And the typical benchmark (especially the cross-platform ones) is not multithreaded.
So unless some breakthrough is achieved relative to multi/single-threading, this machine would probably stand out as very slow in every aspect of benchmarking, except where they tested multitasking extensively.
Originally posted by Zapchud
this pseudo-G5 Powerbook would be a terrible machine for all things involving less than 4 threads. A typical example of this is benchmarks.
Apple is the only personal computer company that can change the way computers work. Changing the Power Mac line to dual-processors kicked in SMP on the Mac platform. Cell-style quad-core PowerBooks would further increase the number of threads, and more importantly the efficiency of threads (lot's of programs unfortunatly have token multithreading, I've seen).
Also, I don't give a rat's arse about benchmarks. Let Apple thread their apps better and only give benchmarks for them if necessary.
Barto
Originally posted by Barto
Apple is the only personal computer company that can change the way computers work.....Also, I don't give a rat's arse about benchmarks. Let Apple thread their apps better and only give benchmarks for them if necessary.
I don't know, all this story is confusing. I believe Zapchud has valid points, though I would like to see a powerbook with the processor discussed here. Certainly, Apple will thread further their applications to take advantage of the new architecture, if this happens. And i would prefer to use such a machine over others, since the architecture seems efficient and potentially more reliable than what is available today in the notebook market. Use of cell-like processors by Apple would mark a new era in personal computing.
however as with any revolutionary advancement in computer design, there will be a period in which migration to new programming model is require
apple and IBM have been porting os x to the new programming model.
however
even with single thread, performance will be acceptable for mobile computer. the computer will feel responsive, and processor intensive applications will probably be quickly ported.
Nr9
Do you have actual inside information or is this all just speculation?
Originally posted by Nr9
i have inside info
Thanks, now that was a quick response.
And you're sure this is for a Powerbook? Seems multithreading and multitasking is advancing rapidly, but for this type of system to compete with the brute force of a 9 to 20 stage piped G5 stills seems far fetched to me.
Oh well, hope your right and this system screams.
this system screams with the right software. it can even be faster than current G5
Originally posted by Nr9
...but it will not be faster than power mac avaiable when it it release.
And therein lies the million dollar question.
The key thing here is responsiveness, and this is where Nr9 really has a point. Anyone who's used a duallie remarks on how silk-smooth the interface is. Go up to four processor cores, and you have a machine that can multitask up a storm and still be instantly responsive. Really, the programming model only becomes a big issue if you decide to run a giant single-threaded beast like the average ported video game.
Engineering is always about tradeoffs, but if you're going to have pervasive tasking and threading - as OS X does - then many hands make light work of it.
Originally posted by Amorph
The only remaining champion of the traditional personal computer architecture is Intel.........This sort of 440-based board is one of the mammals scurrying around under the feet of the mighty dinosaurs........
Not that my opinion is worth anything, but I would like to state that this is the best, most thought provoking post I'ver ever read on AI. And it manages to include: a dinosaur analogy and the absolutely grammatically correct spelling of 'de-emphasize' (go and look!)
Bravo, Amorph!
Yum
Originally posted by Henriok
According to roadmaps I've seen Mojave will be here in 2004, and Orion in 2005.
Arent Orion something with stars (And a Metallica song) and not a desert?
Originally posted by Nr9
even with single thread, performance will be acceptable for mobile computer. the computer will feel responsive, and processor intensive applications will probably be quickly ported.
Riiiight. Photoshop already screeeams!
Originally posted by T'hain Esh Kelch
Arent Orion something with stars (And a Metallica song) and not a desert?
It's a contellation yes. Orion is not a codename for the 7xx-series witch seem to have a desert theme to its codenames (Sahara, Gobi and Mojave), it's a codename in the 4xx-series witch seem to have heaven-related codenames (Aurora and Orion).
And.. Amorph's and Barto's diagrams were very beautiful! I'm not worthy