Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
970MP - Still Coming?
- Thread starter MacRumors
- Start date
- Sort by reaction score
wdlove said:
Yea I got my fingers crossed for WWDC '05 for the single dual-core PowerMacs and Paris '06 for the dual dual-core Xstations.
With regards to the PowerBook G5 coming out - i think that it will be a dual-core G4 variant earlier (i.e WWDC '05/Paris '06) or a full blown G5 variant later on (i.e WWDC '06).
Just my small coppers worth.
Matt
yes, but wrong order 
The new PMGx will be announced at Paris AppleExpo and will probably be dual core for the high-end model. (Latest January 2006)
At the same time Apple will introduce the new PowerBook models based on 7448 for the low profile hypothesis or for the high profile expectation on e600 single core to debute then later one on e600 dual core (WWDC2006). The chance of a PB running with a G5 are rather low... and the benefit close to zero... lower battery lifespan, similar performance, the only difference being 64 bit support. The only problem which was/is present with e600 processor is the Rapid I/O system from Freescale which is new for Apple and has to be implemented. So if in Septembrer the new PB sports a 7448 processor it will mean that Apple is/had trouble to incorporate RapidI/O bus, and is waiting longer to ensure that Freescale will be able to deliver e600 single and dual core according to the roadmap. such hardware work needs to be insured by no other technical problem such as processor issues and/or delay.
the Xstation will be announced at the WWDC for massive parallel processing, it is stil not clear if it will run on 970MP or 970GX, but it will be more than a dual physical CPU system...Cold_Steel said:
The new PMGx will be announced at Paris AppleExpo and will probably be dual core for the high-end model. (Latest January 2006)
At the same time Apple will introduce the new PowerBook models based on 7448 for the low profile hypothesis or for the high profile expectation on e600 single core to debute then later one on e600 dual core (WWDC2006). The chance of a PB running with a G5 are rather low... and the benefit close to zero... lower battery lifespan, similar performance, the only difference being 64 bit support. The only problem which was/is present with e600 processor is the Rapid I/O system from Freescale which is new for Apple and has to be implemented. So if in Septembrer the new PB sports a 7448 processor it will mean that Apple is/had trouble to incorporate RapidI/O bus, and is waiting longer to ensure that Freescale will be able to deliver e600 single and dual core according to the roadmap. such hardware work needs to be insured by no other technical problem such as processor issues and/or delay.
I said this before the PowerMacs were update, and I think it's still valid. If I were at IBM/Apple, and thought how to proceed with the G5, three things comes to my mind:
1. Multicore
2. Intergrated memory-controller
3. Symmetric Multithreading (SMT)
Each of those give nice improvement in performance, with 2. and 3. being relative low-cost things to implement. And IBM knows each of those from their POWER-line of CPU's.
I wont talk that much about multicore (dual-core in practise at this time), since we already know what that is about. But what about the other two?
Integrated mem-controller would significantly boost the effective bandwidth of the system. Yes, G5 has a fast bus, but memory-access eats ALOT of bandwidth, leaving rest of the system (HD's, USB, Firewire etc.) with less bandwidth at it's disposal. Also, it would bring latencies down, since CPU would talk with the RAM directly, instead of through the northbridge.
SMT. Intel has this in P4, and it's called hyperthreading. Basically it would allow the CPU to handle more simultaneous threads, like SMP-systems do. The difference is that while true SMP doubles the number of transistors while giving 70-80% boost in SMP-aware apps/OS'es, SMT give 20-40% boost, while increasing the number of transistors by 5-10%. So it's a good thing to have. And in this case, I'm talking about 2x threads.
So, what would I like to see Apple offer? Currently they offer this:
1x 1.8GHz G5 @ $1499
2x 2.0GHz G5 @ $1999
2x 2.3GHz G5 @ $2499
2x 2.7GHz G5 @ $2999
The current lineup would be replaced by G5 with an integrated mem-controller. I would like to see the 1.8GHz machine dropped. with this machine going between that and the 2x 2.0GHz machine in price:
1x 2.0GHz SMT G5
This machine would be a bit slower than the current 2GHz machine in tasks that require raw CPU-power. But it would be significantly faster in tasks that require lots of mem-bandwidth and/or low latencies. Of course, it would run rings around the current 1.8GHz machine. Price of this machine could be around $1699. It could handle two simultaneous threads.
Next, we could have this machine:
1x 2.4Ghz dualcore SMT G5
This machine could handle four simultaneous threads (two core, two threads per core). Ic would be considerably faster than the current 2x 2.3Ghz machine. Not only would it have more raw MHz, it could handle more simultaneous threads and it would have alot better mem-bandwidth and latencies. And I think it could be a bit cheaper than the current 2.3Ghz machine. Maybe something like $2399.
At the hi-end we would have this:
2x 2.8Ghz dualcore SMT G5
This machine could handle eight simultaneous threads (two CPU's, two cores per CPU, two threads per core). It would have more bandwidth and better latencies than the current 2.7GHz machine, it would have a bit more raw MHz, and it would threads for breakfast! Price of this machine could be around $3299. More than current hi-end, but it would also be alot faster.
The problem with this lineup is that the price-difference between different models would be quite large ($700 and $900, respectively). But then again, the performance-difference would be quite large as well. But even then, there would be a machine for each price-point, and performance would be top-notch.
1. Multicore
2. Intergrated memory-controller
3. Symmetric Multithreading (SMT)
Each of those give nice improvement in performance, with 2. and 3. being relative low-cost things to implement. And IBM knows each of those from their POWER-line of CPU's.
I wont talk that much about multicore (dual-core in practise at this time), since we already know what that is about. But what about the other two?
Integrated mem-controller would significantly boost the effective bandwidth of the system. Yes, G5 has a fast bus, but memory-access eats ALOT of bandwidth, leaving rest of the system (HD's, USB, Firewire etc.) with less bandwidth at it's disposal. Also, it would bring latencies down, since CPU would talk with the RAM directly, instead of through the northbridge.
SMT. Intel has this in P4, and it's called hyperthreading. Basically it would allow the CPU to handle more simultaneous threads, like SMP-systems do. The difference is that while true SMP doubles the number of transistors while giving 70-80% boost in SMP-aware apps/OS'es, SMT give 20-40% boost, while increasing the number of transistors by 5-10%. So it's a good thing to have. And in this case, I'm talking about 2x threads.
So, what would I like to see Apple offer? Currently they offer this:
1x 1.8GHz G5 @ $1499
2x 2.0GHz G5 @ $1999
2x 2.3GHz G5 @ $2499
2x 2.7GHz G5 @ $2999
The current lineup would be replaced by G5 with an integrated mem-controller. I would like to see the 1.8GHz machine dropped. with this machine going between that and the 2x 2.0GHz machine in price:
1x 2.0GHz SMT G5
This machine would be a bit slower than the current 2GHz machine in tasks that require raw CPU-power. But it would be significantly faster in tasks that require lots of mem-bandwidth and/or low latencies. Of course, it would run rings around the current 1.8GHz machine. Price of this machine could be around $1699. It could handle two simultaneous threads.
Next, we could have this machine:
1x 2.4Ghz dualcore SMT G5
This machine could handle four simultaneous threads (two core, two threads per core). Ic would be considerably faster than the current 2x 2.3Ghz machine. Not only would it have more raw MHz, it could handle more simultaneous threads and it would have alot better mem-bandwidth and latencies. And I think it could be a bit cheaper than the current 2.3Ghz machine. Maybe something like $2399.
At the hi-end we would have this:
2x 2.8Ghz dualcore SMT G5
This machine could handle eight simultaneous threads (two CPU's, two cores per CPU, two threads per core). It would have more bandwidth and better latencies than the current 2.7GHz machine, it would have a bit more raw MHz, and it would threads for breakfast! Price of this machine could be around $3299. More than current hi-end, but it would also be alot faster.
The problem with this lineup is that the price-difference between different models would be quite large ($700 and $900, respectively). But then again, the performance-difference would be quite large as well. But even then, there would be a machine for each price-point, and performance would be top-notch.
Evangelion said:
So would this mean that the RAM is physically connected to the CPU, implying separate RAM banks for two CPU chips? Or would some memory controller logic move from the northbridge to the CPU, but the memory access for both chips would still physically pass through the northbridge?
eric67 said:
I've heard MacOSRumors.com speak a lot over the past couple months about "xStation". It will be interesting to see if this comes to fruition. The believe xStation will debut at up to quad 3.2 GHz. And, as you suggest, don't know if this will be 970mp or 970gx.
I've expected dual core PowerMac's at MacWorld Paris, too. It makes sense, considering 2.7 is strategically gapped to 3.0 GHz. (2.0, 2.3, 2.7)
DavidCar said:
What it would mean is that each CPU would have a dedidated RAM-bank, but they could also access the RAM attached to the other CPU. This would be possible since both CPU's would be directly connected to each other, besides being connected to the northbridge (where PCI, USB, FireWire etc. would connect to). In reality that would mean that the memory-bandwidth of the system grows as new CPU's are added to the system.
DavidCar said:
Well, dual-core CPU would not be as expensive to make as two single-core CPU's would be. But they would be considerably more expensive than single-core CPU's would be. And adding SMT and mem-controller there would increase the price of the CPU a bit (but it would reduce costs elsewhere).
But what's important is that the price would be more or less the same as the current 2x 2.3GHz machine, but the performance would be ALOT better. It would have two cores like the 2.3GHz machine (but on single die). But it would also have SMT (something current system does not have), more bandwidth and lower latencies and a bit more raw MHz. And it would still be cheaper than the 2.3Ghz machine!
Will Dual Core PowerMacs or X-Stations Cost More or Less Than Current Line?
I just pulled the trigger on a dual 2.5 refurb to replace my dual 2 refurb cause I can break even on the sale of my dual 2 loaded. In your opinion, Was that a dumb move? You think Steve will drop the dual core bomb June 6?
So you are confusing me. I can't tell if the dual core PowerMacs or X-Stations will cost more or less? Do you think they will REPLACE the existing line soon? It would be highly unusual for Apple to refresh a line right before dumping it soon thereafter.Evangelion said:
I just pulled the trigger on a dual 2.5 refurb to replace my dual 2 refurb cause I can break even on the sale of my dual 2 loaded. In your opinion, Was that a dumb move? You think Steve will drop the dual core bomb June 6?
Evangelion said:
My first reaction is that such an architecture sounds more elaborate than what Apple would be inclined to develop. Do you mean four cores on two chips would require four RAM banks? It sounds like such an architecture could lay the foundation of a modular machine, where more CPU/RAM banks could be added to the initial machine, with a multiport northbridge providing for additional CPUs to connect to each other's RAM banks, (with considerable delay) and to the PCI, USB etc.
Wouldn't that require significant change to the OS to determine which RAM bank some information should reside in? Or would that just be a slight variation of how the OS is handling multi-machine supercomputer configurations?
Multimedia said:So you are confusing me. I can't tell if the dual core PowerMacs or X-Stations will cost more or less? Do you think they will REPLACE the existing line soon? It would be highly unusual for Apple to refresh a line right before dumping it soon thereafter.
I just pulled the trigger on a dual 2.5 refurb to replace my dual 2 refurb cause I can break even on the sale of my dual 2 loaded. In your opinion, Was that a dumb move? You think Steve will drop the dual core bomb June 6?
I'm not sure what would count as a replacement and what doesn't. Current lineup has different vid-cards and CPU's as the original G5-PM's did. Well, they could add in PCI-Express besides dual-core G5's as well.
What I have posted is purely speculation on my part
. I have no information one way or the other regarding G5 and it's future versions. What I have posted is what I would do, if I were in a position to influence things.No, you didn't do anything stupid. Enjoy your new uber-powerful machine
. Yes, sometime in the future Apple will release new machine that will be faster. But that's the way it always works DavidCar said:
No, the integrated mem-controller would be per-CPU, not per-core. So the machine you described would have just two banks of RAM.
The OS would have to be NUMA-aware. And it's doable. Linux was made NUMA-aware in the 2.6-series.
DavidCar said:
A memory bank per processor is only useful in Systems that have hyper-transport(such as Opteron systems).To truly benefit from this design they also require NUMA capability in the kernel. Something that Darwin/MACH doesn't have and probably won't ever have. Heck Darwin/MACH has the worst memory subsystem of any OS on the market today and Apple isn't in any hurry to change it.
Evangelion said:
I doubt that Apple has the vision or expertise to ever gut MACH/Darwin to actually make something useful when it comes to a memory subsystem.
For the Darwin kernel it would require almost a complete rewrite from the ground up.
All this talk about G5s G6s Xstations, well whatever, Apple better have something ready up its sleeve because this is their time to gain market share and they better get the most out of it until MS will release Longhorn... iPods will get you only this far...
Apple proclaimed 2005 to be the year of HD video, they released all this great new software products at NAB and yet they want people to run it on a 2 year old technology...
Apple proclaimed 2005 to be the year of HD video, they released all this great new software products at NAB and yet they want people to run it on a 2 year old technology...
~loserman~ said:
I understand that one of the changes in Tiger is improving the interface to the kernel so that changes to the workings of the kernel would be possible. I don't understand the details, but it would be interesting to know what the possible changes could be.
http://arstechnica.com/reviews/os/macosx-10.4.ars/4
Or specifically, if anyone knows the details, would the changes in the Tiger kernel interface make improvements such as NUMA capability easier?
FWIW, I looked up NUMA, or Non Uniform Memory Access, and found this:
http://lse.sourceforge.net/numa/faq/
...and I thought Apple was into HyperTransport, being part of the HyperTransport Consortium along with IBM:
https://www.macrumors.com/pages/2003/08/20030811144341.shtml
So wouldn't it be no surprise if a 970MP used HyperTransport?
Yes, it would be a big surprise -- almost as big as Ford letting GM rebadge a Lincoln Navigator as a Cadillac.DavidCar said:
The FSB of the 970MP is the elastic-i/o bus, changing it to HT would be a big change -- massive.
HT is just glue between the Northbridge and Southbridge, nothing more right now.
---
Remember, at the time Apple went with the PPC970 and the elastic bus -- HT didn't have a public SMP solution (AMD kept it private) and it would NOT navigate board hops (like our current processor daughtercards) and had a bad trace length limitation (had to be kept short).
RapidIO was the only solution with a board hopping, public SMP cache snooping capability, and could also be used to connect devices on the farthest corners of a motherboard.
Motorola is using RapidIO on their embedded G4 application, RIO isn't too different from HT though -- just uses a different method of sending signals across a trace (I think, been awhile since I read it).
Though we all know IBM delivered using their bus technology instead of Motorola's RapidI/O or AMD's HT.
Sun Baked said:
Thanks for clueing me in about bus specifications.
Earlier, Loserman made the assertion that "A memory bank per processor is only useful in Systems that have hyper-transport(such as Opteron systems)" The context is a consideration of whether a 970MP could have an on die memory controller, which would imply a memory bank per processor, and perhaps enable expandable systems with a processor and memory bank on one or more cards, with NUMA capability for accessing memory across cards.
If ElasticIO doesn't support such a scenario, then that might be an argument against seeing an on die memory controller anytime soon.
You people are dreaming if you think dual core systems are going to cost the same as the current line up or even CHEAPER??? Why don't you take a look at prices for Dual core systems that currently exist (for PCs). 2.5ish 'single' dual core workstations start at about $4000, get that closer to 3ghz and throw in a second dual-coe proc and you should be looking at about 5-6000 (the amd 2ghz+ duals are all over $1000...and thats just the proc alone).
the AMD dual core cpu's are about $900. however, they are only in limited production as AMD wants to phase them in and keep the market open for their other chips. i guess if they produced a lot of them the price would come down. my guess is apple would buy enough of any cpu to make it cost effective for their line of computers. just a thought though.
crpchristian said:
Exactly my thoughts... But then again there are rumors that Apple will release an all new Pro-line which I am guessing will be above Powermacs (performance and pricewise)... But is there a market for such a machine? If there was i am guessing Apple would have released a Quad processor computer a while ago...
KindredMAC said:
The G4s were updated right before WWDC in '03. the dual 1.42 was available mid-March or so (the first PMs 1.25 and 1.42 wih FW800) and the G5 was announced just 2.5 months later at the June meeting. The 1.42 only continued on until July for a total life cycle of 4.5 months on the market.
Also would like to see the ultra high speed Rambus XDR Ram technology used (Toshiba just announced a second generation product) and PCI-X 16 video (now well est. in the PC world) in the new line.