So there are 3 scenarios in which the OS, and anyone trying out a "pre-release" machine, will see a "4 processor" Mac:
1. The machine actually has 4 CPU chips.
2. The machine has two dual-core chips, putting 4 cores on 2 pieces of silicon.
3. The machine has 2 Hyperthreading CPUs.
In any of the above situations, the OS will treat it as 4 processors. So even if MOSR has some inside information, they could easily be mistaken as to the exact nature of the "quad". I have no doubt that Apple, though, would call any of the above machines "quad" for marketing purposes.
Considering bus saturation, diminishing returns on SMP, and lack of pervasive multithreading in many Mac apps, unfortunately the performance probably wouldn't be too different between the above 3 scenarios in most situations. Given that #3 would be far, far cheaper, that's what I'd guess Apple would do.
On the bright side, it will be a lot tougher to saturate the bus on these machines .
MOSR aside, I have thought that the Quad Powermac is a direction that Apple should have taken since the G4 7410 around 1999.
Providing the market with a maxium performance machine is something that Apple has been reluctant to do. On the other hand Apple now has competition that it didn't have in the past.
Quote:
It was at the height of the 500 MHz speed stall that made me think that this was the next logical step for Apple to remain competitive with the Wintel MHz gap. A search for "quad G4" will show that several vendors actually offered 7410 G4s in a four way configuration (mostly for DSP applications).
At the tiem this way the only way such machines could accell. That is as special purpose proecessors. Apple no has just about all of the software pieces in place so that this is no long an issue. Further, the current 970's are far better suited for extensive SMP work than just about nay processor except possibly AMD's Opteron (desktop people, lets not digress with other chips)
Quote:
The DayStar Genesis MP offered a four processor (604e's) box before Apple pulled the plug on the clones. This advancement in the platform, leaving Apple's top end in the dust, may have been instrumental in the clones demise.
But that was then, and this is now. If Apple didn't jump on this concept then (when they really needed it) why would they bother now? More importatantly, would a quad Powermac (xStation?) sell, if so, to whom?
Probally because they will have no choice as the technology is quickly moving to dual core processors doing SMP as the base hardware. Some may not want to admit to that but the reality is that in order for dual core chips to be economical they have to be adopted on a wide scale. I would not be surprised that a year or two after dual cores hit the market all mid level hardware will be 2X SMP. Going to 4X SMP will only involve dropping in another chip.
Not to be underestimated is SMT which will in effect offer another thread of execution. It is very possible that we may see far better SMT support form IBM than has been seen on ohter processors. For some applications this means 4 to 8 logical processors performing nearly as well as 2 or 4.
To whom will these be sold to: --- EVERYONE --- once the software is ironed out and apps that truely benefit hit the market, you will quickly end up seeing a great deal of used hardware on the market. Frankly who would want a SMP machine for the price of a single core machine? Especially if the cost of the machine is not impacted and we are still seeing signifcant core speed ups.
Quote:
I have my own ideas about this (who would buy such an Apple) but would like to hear from other posters about the practicality and possibility of Apple going this direction.
Apple simply has no choice, to remain a player in the market they will need to increase the number of threads being executed at the same time. It will be an absolute requirement in the market in a very short period of time. There are a number of approaches to getting this to happen, you have SMT, Dual Core Processor, 4X SMP and variants where SMT is mixed with Dual Core technology for example. You can also have significant variation with respect to how Dual Core is implemented. IN the end each technology will have its own advantages and disadvantages with Apple adopting what supports its market segment best.
So I expect that it is extremely possible that Apple will have at least 4 logical processors in its machines in very short order with a very good possibility that they may have many more than that. It may be very possible to up that to 8 logical processors with out having a significant impact on the economics of the machines. In fact one should not expect a significant price increase especially if they can implement a Dual Core SMT chip with todays technology.
Look at it this way, GPU's which are now often bigger than the CPU's in the same machine are still affordable. Depending on the exact possiblity that Apple/IBM pursues the resulting Dual Core chip will still be smaller (transistor count) than many of todays GPU's. It may even be smaller (physical size) than the original 970. One should not get to excited about Dual Core until we actually see the real thing, there are many possibilities here (just like SMT) as to hw well the processor will perform.
Quote:
Any thoughts?
Always one more! Look closely at what AMD is doing, they currently have some of the best technology on the market with respect to multi processing. They also are much more open to what is happening with their product line.
Considering bus saturation, diminishing returns on SMP, and lack of pervasive multithreading in many Mac apps, unfortunately the performance probably wouldn't be too different between the above 3 scenarios in most situations. Given that #3 would be far, far cheaper, that's what I'd guess Apple would do.
While I agree with the first part of your comment I do wonder where this negativity comes from.
Apple currently has the best desktop environment for multithreading and multiprocessing on the market. While there may be technical aspect of Apples kernel that need improvement, it is very easy for one to mis the fact that the much of the OS is working towards being multiprocessing aware and is currently doing very well agianst its desktop competition.
In the applications where it really matters they are very much multithreaded. Where it doesn't matter they aren't obviously. In between we do have a bunch of Apps that could benefit but haven't to date. It is a stretch to call the lack of threaded applications pervasive, for some apps the pay off will always be hard to justify.
In any event it is foolish to focus on single application performance in a multiprocessing system. Better to focus on the over all performance of the machine when running a variety of tasks at once. This is what multiprocessing does well. As far as diminishing returns on SMP that is a limit that is far off for Apple. It is very possible to get a 4X SMP machine in an ATX box that is a very storng performer and nothing to see 32 or 64 processor in bigger machines. Most of the limits that one see with respect to desktop SMP with more than 2 processors are the result of software support.
Programmer and I hashed out the most likely course a while ago, and I still think it's right: Since motherboards that accomodate more than two CPUs start becoming really complex and expensive really quickly, what you'll see is 1 or 2 CPUs which have more and more cores, and SMT as an added boost. There's a scaling limit there, too, because lots of (full, traditional) cores on a single bus run into the same problem that plagued G4 systems with more than two CPUs on the same bus: There's not enough bandwidth to go around.
I suspect that Cell is intended to be a way around this problem.
While a quad system would certainly be nice, I don't really see it happening (except perhaps 2 dual-core processors). The people in the business of needing a lot of processors aren't going to rely on a single user workstation, but an array of Xserve cluster nodes sitting in a closet.
Apple hits the Dual 3ghz mark with the next Powermac refresh. This will be a 97x G5 based off of the POWER5 architecture. It will not be dual core but it will have say 5 more pipeline stages(roughly 20) and support low-k dielectric tech. It will have SMT so we will have our "functional" Quad. Apple will ride this SMT enabled system for a year or year and a half at most before IBM movies to 65nm in which they will move to dual core processors then. The question is whether or not they would keep the SMT on each core. If they did that would totally rock. How's an Octomac sound to you?
What I see is a balancing of todays technology. SMT gives Apple/IBM a reprieve from being forced to go dual core early. SMT allows IBM to increase the pipelines thus hitting 3Ghz easier without taking a huge hit on Instructions Per Cycle(IPC) because SMT will keep the pipelines full with data. SMT is the cheaper option right now that being that it requires much less transistor space than a whole "nother core. That space savings can be put to good use by adding an ondie memory controller. Then when a majority of Mac apps are multi processor aware we move to the next process shrink 65nm which then gives us die space to add those new cores. We're only at 2007 at this time frame. Nice logical progression.
While I agree with the first part of your comment I do wonder where this negativity comes from.
The current dual-processor 2GHz machine I have uses a dual-channel, 1GHz, 64-bit pipe from each processor to the memory controller. It then uses a 800MHz, single channel DDR pipe from the memory controller to RAM. I don't see any way to make a sub-$5000 machine with fast enough RAM to fully feed the processors in the current design, let alone four 3GHz processors. That's what I was referring to by bus saturation.
By pervasive multithreading, I was talking on the BeOS-level of things. For real professionals, apps like Photoshop and FCP have the threads. But for home uses, such as Excel, America's Army, etc., it's not going to help at all. While BeOS had a thread per window and a main app thread, Apple ties the graphics into the main thread and thus by default produces single-threaded apps. That's why BeOS was so insanely responsive on a PowerPC 603e, while MacOS is still sluggish resizing iTunes windows on a G5.
Basically, multithreading seems like "enough" of a boost to be noticable, and allow the UI responsiveness of more processors, without actually saturating the bus significantly more. (ie. It's like adding more low-end torque to a car while leaving the horsepower almost the same.)
And it's cheaper. WAY cheaper. Apple, being a company that likes to make a profit, would be able to make higher margins on the simpler design. Thus, I think the likelihood here is that we'll get my option 3 (hyperthreading) in the next release, option 2 (2 chip, dual-core each) in 2006-2007, and option 1 (four processors) never.
Apples strategy has but to forgo on threading everything and simply evolve the Quartz model. With Tiger we are likely to see more improvement in todays Quartz Compositor along with the the intro of Quartz 2D Extreme which should clear up the nagging issues with slow window redraw.
10.4.x or 10.5 will bring us Resolution Independence when its ready.
As for memory controllers I agree with Booga. Even todays model isn't as efficient as I'd like. Adding 4 actuall CPUs means you have have to add more memory banks also. Now we're getting expensive.
What I'd like to see is Apple move to SMT for more efficiency...let developers work on supporting SMP and threading in their apps a bit more..then we can move to dual core and hopefully ondie memory controllers as well. By that time we'll have DDR2 800-1000. That gives us 12,800 and 16,000GBps throughput available. Link each CPU with HT 2.0 which supports up to 22GBps throughput and you have a system that can keep up with enormous amounts of bandwidth.
Do you really think they will need to lengthen the pipeline that much? My impression of the 970 is that it and Power4 share in a designe that is IBM's stepping stone to SMT. Power 5 being a minor tweak of the 970/Power 4 to fully realize that capability.
Ok so maybe there is a bit of Rose tinting in the glasses but the little bit of information I've seen on IBM's web site seems to support that the Power 5 is an evolution of the power architecture.
SMT could be a stop gap measure for Apple/IBM but it is a measure that is not going to be a predictable when it comes to performance realization. Then agian dual cores processors could have their own problems. I see Apple going dual core on their midrange machines and to 4X processors on their higher end machines.
Dave
Quote:
Originally posted by hmurchison
SMT gives Apple/IBM a reprieve from being forced to go dual core early. SMT allows IBM to increase the pipelines thus hitting 3Ghz easier without taking a huge hit on Instructions Per Cycle(IPC) because SMT will keep the pipelines full with data. SMT is the cheaper option right now that being that it requires much less transistor space than a whole "nother core. That space savings can be put to good use by adding an ondie memory controller. Then when a majority of Mac apps are multi processor aware we move to the next process shrink 65nm which then gives us die space to add those new cores. We're only at 2007 at this time frame. Nice logical progression.
The current dual-processor 2GHz machine I have uses a dual-channel, 1GHz, 64-bit pipe from each processor to the memory controller. It then uses a 800MHz, single channel DDR pipe from the memory controller to RAM. I don't see any way to make a sub-$5000 machine with fast enough RAM to fully feed the processors in the current design, let alone four 3GHz processors. That's what I was referring to by bus saturation.
There is no reason why Apple has to follow the blue print of the current PowerMac G5. There are alternatives some of which are available from AMD for the desktop market. Your concerns are warranted but there is no reason to try to extend the current model.
[qoute]
By pervasive multithreading, I was talking on the BeOS-level of things. For real professionals, apps like Photoshop and FCP have the threads. But for home uses, such as Excel, America's Army, etc., it's not going to help at all. While BeOS had a thread per window and a main app thread, Apple ties the graphics into the main thread and thus by default produces single-threaded apps. That's why BeOS was so insanely responsive on a PowerPC 603e, while MacOS is still sluggish resizing iTunes windows on a G5.
[/quote]
We are still working on a multiprocessing system so on a process basis you are still taking advantage of the hardware. To a certian extent though even single thread applications are advantaged by those parts of the OS that Apple has gotten around to threading. I don't deny that there is room for improvement, what OS doesn't have room for improvement. On the other hand developers are not stuck with the default araingement.
Quote:
Basically, multithreading seems like "enough" of a boost to be noticable, and allow the UI responsiveness of more processors, without actually saturating the bus significantly more. (ie. It's like adding more low-end torque to a car while leaving the horsepower almost the same.)
I look at multithreading as a tool that developers can take advantage of. Sure for some that menas an enhanced UI. For others the UI means really nothing, what they want from multithreading is the ability to get more work done by exploiting the resources of the machine.
Quote:
And it's cheaper. WAY cheaper. Apple, being a company that likes to make a profit, would be able to make higher margins on the simpler design. Thus, I think the likelihood here is that we'll get my option 3 (hyperthreading) in the next release, option 2 (2 chip, dual-core each) in 2006-2007, and option 1 (four processors) never.
I most certianly agree that SMT is the strong play for the next series of machines. On the other hand I do not believe that Dual Core is that far off. Dual Core may come with SMT it is hard to say, even Apples plans may be in question as the competive environment has changed dramatically. Both Intel and AMD have indicated that Dual Core processors are a extremely high priority, with Intel literally crashing its development process. Both of these guys are talking early 2005, from a marketing standpoint Apple/IBM may not have much of a choice but to go Dual Core.
I still believe that we are on the verge of a minor revolution in desktop computer technology. In part that will mean the migration across the board to SMP in some form or another. Along with this will be dramtically faster busses, co-processors, and memory systems. As you may geuss I don't have crystal ball, much of this is public information. What I feel many fail to realize, is that in a years time we will have dramtically faster machines on our desks that may actually be less expensive than todays.
Yes software will spend time catching up but isn't that always the case.
Dave I agree with you I think the race is to dual core. I'm just being overly pessimistic about Apple being aggressive in going dc.
We still don't have the seperation needed betwen the iMacs and Powermacs. We should never even have to worry about the iMac stepping on the Powermacs toes unless the Powermacs don't contain enough Pro features for people to see the added value.
Threading is a requirement for any modern OS! It is not something that Apple can ignore and expect to be taken in a serious manner. Not that Apple is takend seriously now but if they want to gain credibility this is one thing that has to be implemented and done soundly.
Quote:
Apples strategy has but to forgo on threading everything and simply evolve the Quartz model. With Tiger we are likely to see more improvement in todays Quartz Compositor along with the the intro of Quartz 2D Extreme which should clear up the nagging issues with slow window redraw.
While I can't say what Apples strategy is (I'm not part of the development community) it is clear and has been for some time that Apple takes SMP seriously. Threading can leverage that commitment. As to window redraws and Q 2D E, using the graphics chip only speeds up the drawing process, you still have to have an application that is responsive enough to tell the system to draw whatever. By the way multithreading is not a solution to the problem either, it is a tool that the developers can use in there applications but there is still a responsibility on the developers part. I do not see that going away but maybe a few developers can chime in here.
Quote:
10.4.x or 10.5 will bring us Resolution Independence when its ready.
As for memory controllers I agree with Booga. Even todays model isn't as efficient as I'd like. Adding 4 actuall CPUs means you have have to add more memory banks also. Now we're getting expensive.
Really we need to cut the baloney here. There are already systems on the market that implement 4 processors and do so in an economical manner. Yeah these are AMD systems and the 4 processor mother boards cost a bit more, but they are not unreasonable at all. Apple could follow this path or they could take a different approach with a unified memory system. Each path has its advantages and disadvantages and frankly which is better for Apple probally depends on what work loads they will be targeting with the machine.
Besides there are a whole bunch of people out there who would love to see Apple implement more memory banks in its current machines. Speed may be an issue for some but there are many who simply could use a very large address space.
Quote:
What I'd like to see is Apple move to SMT for more efficiency...let developers work on supporting SMP and threading in their apps a bit more..then we can move to dual core and hopefully ondie memory controllers as well. By that time we'll have DDR2 800-1000. That gives us 12,800 and 16,000GBps throughput available. Link each CPU with HT 2.0 which supports up to 22GBps throughput and you have a system that can keep up with enormous amounts of bandwidth.
Sure things will improve in the future, it is the very near future that has me interested. I suspect that Dual Core will happen very soon maybe with SMT. The issue isn't so much what Apple wants but where the competition is going. Both Intel and AMD expect to have Dual Core SMP systems available by early 2005, I suspect that Apple would love to beat both of those guys to the punch. Not to be excluded here is IBM who would possibly like to obtain a little credibility as a semiconductor provider. If nothing else the coming year will be interesting.
...let developers work on supporting SMP and threading in their apps a bit more...
I think that you will see this when the capability is available in the consumer and laptop lines and there is an easy path to build the threading in doing cross platform development. As to dual cores, IBM already has them working on the Power4 and Power5, so they have plenty of experience with designing them. I wouldn't be suprised to see a 9XX family dual core chip out before the end of 2005.
I would be surprised if Apple does not have a 2 chip dual core 2-way SMT machine by the end of 2006, most likely built on a 65nm process. It will be a POWER4 derived architecture, possibly with extended pipelines. My guess is that this top end processor will not have an on-chip memory controller, but that could go either way (a low end / portable SoC processor is more likely). The top end processor will probably have 2+ MB L2 cache per chip.
On the way there we'll see a single core 2-way SMT processor early-to-mid 2005, starting on 90nm and moving to 65nm in 6-9 months. That's probably IBM's plan... if they run into troubles like they did with 90nm then we could see another 6-9 month slip. If they slip like that then the dual core variant would likely arrive late 2006... if they don't it could arrive earlier, but still in 2006.
Consider that a 2-way SMT 970 w/ 1 MB L2 would be ~100 million transistors (very doable on 90 nm). A dual core version with double the cache would therefore be ~200 million which is easily doable on a 65nm process. I'm actually being rather conservative here -- they could probably jam a bunch more stuff onto this chip and still keep it economical and yielding well. This may include improvements to the integer and AltiVec units to address some of the issues seen in the 970. Enhancements could also come in the form of Cell ALUs, which are probably essentially autonomous vector units not unlike what is found in a modern GPU. So far there is no sign that Apple is at all involved in the Sony, Toshiba, IBM initiative... but if it is POWER based and they deliver as promised then I could see Apple wanting in. This is exactly the kind of hardware enhancement that Apple's OS/hardware/iApp integration would be good at leveraging. Imagine CoreImage and CoreAudio running on 8 3+ GHz vector processors, leaving the main CPU free for everything else. Whoa.
Clock rate is where things are likely to break from the past. Even with a pipeline extension I'm doubtful we'll see much better than 3 GHz; 4, tops, by the start of 2007. Speeds just aren't going to scale the same way. That 10 GHz that was on an old IBM roadmap is likely to remain a pipedream. We might see a billion transistors on a chip by 2010 (imagine: 10 2-way SMT cores!!!), but I'm not nearly as optimistic about the clock rates.
...Consider that a 2-way SMT 970 w/ 1 MB L2 would be ~100 million transistors (very doable on 90 nm). A dual core version with double the cache would therefore be ~200 million which is easily doable on a 65nm process...
Correct me if I am wrong but in a dual core chip isn't there circutry (transistors) that are common between the two cores? Therefore instead of doubling the count of a single core chip it would be slightly less, say 1.8 X the transistors or ~180 million for the dual core (figure pulled off of the top of my head).
Correct me if I am wrong but in a dual core chip isn't there circutry (transistors) that are common between the two cores? Therefore instead of doubling the count of a single core chip it would be slightly less, say 1.8 X the transistors or ~180 million for the dual core (figure pulled off of the top of my head).
Actually, you have to see a dual core chip as two chips on the same piece of silicium. Which is good because it enables fast interconnexion and all, but does not reduce the number of transistors.
The Power 5 chip, also used in Apple Computer Inc.'s higher-end Macintosh (news - web sites) computers, has 276 million transistors and is made using 0.13 micron copper wiring and silicon-on-insulator technologies, which boost performance. It also has a dual-core design, which is akin to two chips in one microprocessor.
Dual core Power 5's are coming! Is this a leak or a misprint? This comes from an IBM Press Release concerning their eServer line. They would know right!
Dual core Power 5's are coming! Is this a leak or a misprint? This comes from an IBM Press Release concerning their eServer line. They would know right!
...
The latest eServer line, which will be available globally on August 31, uses from two to 16 Power 5 processors per server computer and uses as few as one-fourth the number of chips as comparable machines from rival vendors, Armonk, New York-based IBM said in a statement.
For a dual core version of the current chip with SMT and 1 meg L2 cache you're looking at around 125 million transistors compared to the current 58 million. Note though you're doing a bit more than just multicoring it in that example.
Adding a second core and changing nothing else would take it to around 90 million.
Comments
Originally posted by Booga
So there are 3 scenarios in which the OS, and anyone trying out a "pre-release" machine, will see a "4 processor" Mac:
1. The machine actually has 4 CPU chips.
2. The machine has two dual-core chips, putting 4 cores on 2 pieces of silicon.
3. The machine has 2 Hyperthreading CPUs.
In any of the above situations, the OS will treat it as 4 processors. So even if MOSR has some inside information, they could easily be mistaken as to the exact nature of the "quad". I have no doubt that Apple, though, would call any of the above machines "quad" for marketing purposes.
Considering bus saturation, diminishing returns on SMP, and lack of pervasive multithreading in many Mac apps, unfortunately the performance probably wouldn't be too different between the above 3 scenarios in most situations. Given that #3 would be far, far cheaper, that's what I'd guess Apple would do.
On the bright side, it will be a lot tougher to saturate the bus on these machines
Originally posted by Aphelion
MOSR aside, I have thought that the Quad Powermac is a direction that Apple should have taken since the G4 7410 around 1999.
Providing the market with a maxium performance machine is something that Apple has been reluctant to do. On the other hand Apple now has competition that it didn't have in the past.
It was at the height of the 500 MHz speed stall that made me think that this was the next logical step for Apple to remain competitive with the Wintel MHz gap. A search for "quad G4" will show that several vendors actually offered 7410 G4s in a four way configuration (mostly for DSP applications).
At the tiem this way the only way such machines could accell. That is as special purpose proecessors. Apple no has just about all of the software pieces in place so that this is no long an issue. Further, the current 970's are far better suited for extensive SMP work than just about nay processor except possibly AMD's Opteron (desktop people, lets not digress with other chips)
The DayStar Genesis MP offered a four processor (604e's) box before Apple pulled the plug on the clones. This advancement in the platform, leaving Apple's top end in the dust, may have been instrumental in the clones demise.
But that was then, and this is now. If Apple didn't jump on this concept then (when they really needed it) why would they bother now? More importatantly, would a quad Powermac (xStation?) sell, if so, to whom?
Probally because they will have no choice as the technology is quickly moving to dual core processors doing SMP as the base hardware. Some may not want to admit to that but the reality is that in order for dual core chips to be economical they have to be adopted on a wide scale. I would not be surprised that a year or two after dual cores hit the market all mid level hardware will be 2X SMP. Going to 4X SMP will only involve dropping in another chip.
Not to be underestimated is SMT which will in effect offer another thread of execution. It is very possible that we may see far better SMT support form IBM than has been seen on ohter processors. For some applications this means 4 to 8 logical processors performing nearly as well as 2 or 4.
To whom will these be sold to: --- EVERYONE --- once the software is ironed out and apps that truely benefit hit the market, you will quickly end up seeing a great deal of used hardware on the market. Frankly who would want a SMP machine for the price of a single core machine? Especially if the cost of the machine is not impacted and we are still seeing signifcant core speed ups.
I have my own ideas about this (who would buy such an Apple) but would like to hear from other posters about the practicality and possibility of Apple going this direction.
Apple simply has no choice, to remain a player in the market they will need to increase the number of threads being executed at the same time. It will be an absolute requirement in the market in a very short period of time. There are a number of approaches to getting this to happen, you have SMT, Dual Core Processor, 4X SMP and variants where SMT is mixed with Dual Core technology for example. You can also have significant variation with respect to how Dual Core is implemented. IN the end each technology will have its own advantages and disadvantages with Apple adopting what supports its market segment best.
So I expect that it is extremely possible that Apple will have at least 4 logical processors in its machines in very short order with a very good possibility that they may have many more than that. It may be very possible to up that to 8 logical processors with out having a significant impact on the economics of the machines. In fact one should not expect a significant price increase especially if they can implement a Dual Core SMT chip with todays technology.
Look at it this way, GPU's which are now often bigger than the CPU's in the same machine are still affordable. Depending on the exact possiblity that Apple/IBM pursues the resulting Dual Core chip will still be smaller (transistor count) than many of todays GPU's. It may even be smaller (physical size) than the original 970. One should not get to excited about Dual Core until we actually see the real thing, there are many possibilities here (just like SMT) as to hw well the processor will perform.
Any thoughts?
Always one more! Look closely at what AMD is doing, they currently have some of the best technology on the market with respect to multi processing. They also are much more open to what is happening with their product line.
Thanks
Dave
...
Originally posted by Booga
Considering bus saturation, diminishing returns on SMP, and lack of pervasive multithreading in many Mac apps, unfortunately the performance probably wouldn't be too different between the above 3 scenarios in most situations. Given that #3 would be far, far cheaper, that's what I'd guess Apple would do.
While I agree with the first part of your comment I do wonder where this negativity comes from.
Apple currently has the best desktop environment for multithreading and multiprocessing on the market. While there may be technical aspect of Apples kernel that need improvement, it is very easy for one to mis the fact that the much of the OS is working towards being multiprocessing aware and is currently doing very well agianst its desktop competition.
In the applications where it really matters they are very much multithreaded. Where it doesn't matter they aren't obviously. In between we do have a bunch of Apps that could benefit but haven't to date. It is a stretch to call the lack of threaded applications pervasive, for some apps the pay off will always be hard to justify.
In any event it is foolish to focus on single application performance in a multiprocessing system. Better to focus on the over all performance of the machine when running a variety of tasks at once. This is what multiprocessing does well. As far as diminishing returns on SMP that is a limit that is far off for Apple. It is very possible to get a 4X SMP machine in an ATX box that is a very storng performer and nothing to see 32 or 64 processor in bigger machines. Most of the limits that one see with respect to desktop SMP with more than 2 processors are the result of software support.
I suspect that Cell is intended to be a way around this problem.
Apple hits the Dual 3ghz mark with the next Powermac refresh. This will be a 97x G5 based off of the POWER5 architecture. It will not be dual core but it will have say 5 more pipeline stages(roughly 20) and support low-k dielectric tech. It will have SMT so we will have our "functional" Quad. Apple will ride this SMT enabled system for a year or year and a half at most before IBM movies to 65nm in which they will move to dual core processors then. The question is whether or not they would keep the SMT on each core. If they did that would totally rock. How's an Octomac sound to you?
What I see is a balancing of todays technology. SMT gives Apple/IBM a reprieve from being forced to go dual core early. SMT allows IBM to increase the pipelines thus hitting 3Ghz easier without taking a huge hit on Instructions Per Cycle(IPC) because SMT will keep the pipelines full with data. SMT is the cheaper option right now that being that it requires much less transistor space than a whole "nother core. That space savings can be put to good use by adding an ondie memory controller. Then when a majority of Mac apps are multi processor aware we move to the next process shrink 65nm which then gives us die space to add those new cores. We're only at 2007 at this time frame. Nice logical progression.
Originally posted by wizard69
While I agree with the first part of your comment I do wonder where this negativity comes from.
The current dual-processor 2GHz machine I have uses a dual-channel, 1GHz, 64-bit pipe from each processor to the memory controller. It then uses a 800MHz, single channel DDR pipe from the memory controller to RAM. I don't see any way to make a sub-$5000 machine with fast enough RAM to fully feed the processors in the current design, let alone four 3GHz processors. That's what I was referring to by bus saturation.
By pervasive multithreading, I was talking on the BeOS-level of things. For real professionals, apps like Photoshop and FCP have the threads. But for home uses, such as Excel, America's Army, etc., it's not going to help at all. While BeOS had a thread per window and a main app thread, Apple ties the graphics into the main thread and thus by default produces single-threaded apps. That's why BeOS was so insanely responsive on a PowerPC 603e, while MacOS is still sluggish resizing iTunes windows on a G5.
Basically, multithreading seems like "enough" of a boost to be noticable, and allow the UI responsiveness of more processors, without actually saturating the bus significantly more. (ie. It's like adding more low-end torque to a car while leaving the horsepower almost the same.)
And it's cheaper. WAY cheaper. Apple, being a company that likes to make a profit, would be able to make higher margins on the simpler design. Thus, I think the likelihood here is that we'll get my option 3 (hyperthreading) in the next release, option 2 (2 chip, dual-core each) in 2006-2007, and option 1 (four processors) never.
Apples strategy has but to forgo on threading everything and simply evolve the Quartz model. With Tiger we are likely to see more improvement in todays Quartz Compositor along with the the intro of Quartz 2D Extreme which should clear up the nagging issues with slow window redraw.
10.4.x or 10.5 will bring us Resolution Independence when its ready.
As for memory controllers I agree with Booga. Even todays model isn't as efficient as I'd like. Adding 4 actuall CPUs means you have have to add more memory banks also. Now we're getting expensive.
What I'd like to see is Apple move to SMT for more efficiency...let developers work on supporting SMP and threading in their apps a bit more..then we can move to dual core and hopefully ondie memory controllers as well. By that time we'll have DDR2 800-1000. That gives us 12,800 and 16,000GBps throughput available. Link each CPU with HT 2.0 which supports up to 22GBps throughput and you have a system that can keep up with enormous amounts of bandwidth.
Ok so maybe there is a bit of Rose tinting in the glasses but the little bit of information I've seen on IBM's web site seems to support that the Power 5 is an evolution of the power architecture.
SMT could be a stop gap measure for Apple/IBM but it is a measure that is not going to be a predictable when it comes to performance realization. Then agian dual cores processors could have their own problems. I see Apple going dual core on their midrange machines and to 4X processors on their higher end machines.
Dave
Originally posted by hmurchison
SMT gives Apple/IBM a reprieve from being forced to go dual core early. SMT allows IBM to increase the pipelines thus hitting 3Ghz easier without taking a huge hit on Instructions Per Cycle(IPC) because SMT will keep the pipelines full with data. SMT is the cheaper option right now that being that it requires much less transistor space than a whole "nother core. That space savings can be put to good use by adding an ondie memory controller. Then when a majority of Mac apps are multi processor aware we move to the next process shrink 65nm which then gives us die space to add those new cores. We're only at 2007 at this time frame. Nice logical progression.
Originally posted by Booga
The current dual-processor 2GHz machine I have uses a dual-channel, 1GHz, 64-bit pipe from each processor to the memory controller. It then uses a 800MHz, single channel DDR pipe from the memory controller to RAM. I don't see any way to make a sub-$5000 machine with fast enough RAM to fully feed the processors in the current design, let alone four 3GHz processors. That's what I was referring to by bus saturation.
There is no reason why Apple has to follow the blue print of the current PowerMac G5. There are alternatives some of which are available from AMD for the desktop market. Your concerns are warranted but there is no reason to try to extend the current model.
[qoute]
By pervasive multithreading, I was talking on the BeOS-level of things. For real professionals, apps like Photoshop and FCP have the threads. But for home uses, such as Excel, America's Army, etc., it's not going to help at all. While BeOS had a thread per window and a main app thread, Apple ties the graphics into the main thread and thus by default produces single-threaded apps. That's why BeOS was so insanely responsive on a PowerPC 603e, while MacOS is still sluggish resizing iTunes windows on a G5.
[/quote]
We are still working on a multiprocessing system so on a process basis you are still taking advantage of the hardware. To a certian extent though even single thread applications are advantaged by those parts of the OS that Apple has gotten around to threading. I don't deny that there is room for improvement, what OS doesn't have room for improvement. On the other hand developers are not stuck with the default araingement.
Basically, multithreading seems like "enough" of a boost to be noticable, and allow the UI responsiveness of more processors, without actually saturating the bus significantly more. (ie. It's like adding more low-end torque to a car while leaving the horsepower almost the same.)
I look at multithreading as a tool that developers can take advantage of. Sure for some that menas an enhanced UI. For others the UI means really nothing, what they want from multithreading is the ability to get more work done by exploiting the resources of the machine.
And it's cheaper. WAY cheaper. Apple, being a company that likes to make a profit, would be able to make higher margins on the simpler design. Thus, I think the likelihood here is that we'll get my option 3 (hyperthreading) in the next release, option 2 (2 chip, dual-core each) in 2006-2007, and option 1 (four processors) never.
I most certianly agree that SMT is the strong play for the next series of machines. On the other hand I do not believe that Dual Core is that far off. Dual Core may come with SMT it is hard to say, even Apples plans may be in question as the competive environment has changed dramatically. Both Intel and AMD have indicated that Dual Core processors are a extremely high priority, with Intel literally crashing its development process. Both of these guys are talking early 2005, from a marketing standpoint Apple/IBM may not have much of a choice but to go Dual Core.
I still believe that we are on the verge of a minor revolution in desktop computer technology. In part that will mean the migration across the board to SMP in some form or another. Along with this will be dramtically faster busses, co-processors, and memory systems. As you may geuss I don't have crystal ball, much of this is public information. What I feel many fail to realize, is that in a years time we will have dramtically faster machines on our desks that may actually be less expensive than todays.
Yes software will spend time catching up but isn't that always the case.
Dave
We still don't have the seperation needed betwen the iMacs and Powermacs. We should never even have to worry about the iMac stepping on the Powermacs toes unless the Powermacs don't contain enough Pro features for people to see the added value.
Originally posted by hmurchison
Threading is not going to be necessary for OSX.
Threading is a requirement for any modern OS! It is not something that Apple can ignore and expect to be taken in a serious manner. Not that Apple is takend seriously now but if they want to gain credibility this is one thing that has to be implemented and done soundly.
Apples strategy has but to forgo on threading everything and simply evolve the Quartz model. With Tiger we are likely to see more improvement in todays Quartz Compositor along with the the intro of Quartz 2D Extreme which should clear up the nagging issues with slow window redraw.
While I can't say what Apples strategy is (I'm not part of the development community) it is clear and has been for some time that Apple takes SMP seriously. Threading can leverage that commitment. As to window redraws and Q 2D E, using the graphics chip only speeds up the drawing process, you still have to have an application that is responsive enough to tell the system to draw whatever. By the way multithreading is not a solution to the problem either, it is a tool that the developers can use in there applications but there is still a responsibility on the developers part. I do not see that going away but maybe a few developers can chime in here.
10.4.x or 10.5 will bring us Resolution Independence when its ready.
As for memory controllers I agree with Booga. Even todays model isn't as efficient as I'd like. Adding 4 actuall CPUs means you have have to add more memory banks also. Now we're getting expensive.
Really we need to cut the baloney here. There are already systems on the market that implement 4 processors and do so in an economical manner. Yeah these are AMD systems and the 4 processor mother boards cost a bit more, but they are not unreasonable at all. Apple could follow this path or they could take a different approach with a unified memory system. Each path has its advantages and disadvantages and frankly which is better for Apple probally depends on what work loads they will be targeting with the machine.
Besides there are a whole bunch of people out there who would love to see Apple implement more memory banks in its current machines. Speed may be an issue for some but there are many who simply could use a very large address space.
What I'd like to see is Apple move to SMT for more efficiency...let developers work on supporting SMP and threading in their apps a bit more..then we can move to dual core and hopefully ondie memory controllers as well. By that time we'll have DDR2 800-1000. That gives us 12,800 and 16,000GBps throughput available. Link each CPU with HT 2.0 which supports up to 22GBps throughput and you have a system that can keep up with enormous amounts of bandwidth.
Sure things will improve in the future, it is the very near future that has me interested. I suspect that Dual Core will happen very soon maybe with SMT. The issue isn't so much what Apple wants but where the competition is going. Both Intel and AMD expect to have Dual Core SMP systems available by early 2005, I suspect that Apple would love to beat both of those guys to the punch. Not to be excluded here is IBM who would possibly like to obtain a little credibility as a semiconductor provider. If nothing else the coming year will be interesting.
Thanks
Dave
Originally posted by hmurchison
...let developers work on supporting SMP and threading in their apps a bit more...
I think that you will see this when the capability is available in the consumer and laptop lines and there is an easy path to build the threading in doing cross platform development. As to dual cores, IBM already has them working on the Power4 and Power5, so they have plenty of experience with designing them. I wouldn't be suprised to see a 9XX family dual core chip out before the end of 2005.
On the way there we'll see a single core 2-way SMT processor early-to-mid 2005, starting on 90nm and moving to 65nm in 6-9 months. That's probably IBM's plan... if they run into troubles like they did with 90nm then we could see another 6-9 month slip. If they slip like that then the dual core variant would likely arrive late 2006... if they don't it could arrive earlier, but still in 2006.
Consider that a 2-way SMT 970 w/ 1 MB L2 would be ~100 million transistors (very doable on 90 nm). A dual core version with double the cache would therefore be ~200 million which is easily doable on a 65nm process. I'm actually being rather conservative here -- they could probably jam a bunch more stuff onto this chip and still keep it economical and yielding well. This may include improvements to the integer and AltiVec units to address some of the issues seen in the 970. Enhancements could also come in the form of Cell ALUs, which are probably essentially autonomous vector units not unlike what is found in a modern GPU. So far there is no sign that Apple is at all involved in the Sony, Toshiba, IBM initiative... but if it is POWER based and they deliver as promised then I could see Apple wanting in. This is exactly the kind of hardware enhancement that Apple's OS/hardware/iApp integration would be good at leveraging. Imagine CoreImage and CoreAudio running on 8 3+ GHz vector processors, leaving the main CPU free for everything else. Whoa.
Clock rate is where things are likely to break from the past. Even with a pipeline extension I'm doubtful we'll see much better than 3 GHz; 4, tops, by the start of 2007. Speeds just aren't going to scale the same way. That 10 GHz that was on an old IBM roadmap is likely to remain a pipedream. We might see a billion transistors on a chip by 2010 (imagine: 10 2-way SMT cores!!!), but I'm not nearly as optimistic about the clock rates.
Multi-threaded is the future, get used to it.
Originally posted by Programmer
...Consider that a 2-way SMT 970 w/ 1 MB L2 would be ~100 million transistors (very doable on 90 nm). A dual core version with double the cache would therefore be ~200 million which is easily doable on a 65nm process...
Correct me if I am wrong but in a dual core chip isn't there circutry (transistors) that are common between the two cores? Therefore instead of doubling the count of a single core chip it would be slightly less, say 1.8 X the transistors or ~180 million for the dual core (figure pulled off of the top of my head).
Originally posted by @homenow
Correct me if I am wrong but in a dual core chip isn't there circutry (transistors) that are common between the two cores? Therefore instead of doubling the count of a single core chip it would be slightly less, say 1.8 X the transistors or ~180 million for the dual core (figure pulled off of the top of my head).
Actually, you have to see a dual core chip as two chips on the same piece of silicium. Which is good because it enables fast interconnexion and all, but does not reduce the number of transistors.
Someone correct me if I'm wrong!
Reuters
Yahoo News
The Power 5 chip, also used in Apple Computer Inc.'s higher-end Macintosh (news - web sites) computers, has 276 million transistors and is made using 0.13 micron copper wiring and silicon-on-insulator technologies, which boost performance. It also has a dual-core design, which is akin to two chips in one microprocessor.
Dual core Power 5's are coming! Is this a leak or a misprint? This comes from an IBM Press Release concerning their eServer line. They would know right!
...
Originally posted by Aphelion
HOT OFF THE PRESSES ~ APPLE USES POWER 5 CHIP !!!
Reuters
Yahoo News
Dual core Power 5's are coming! Is this a leak or a misprint? This comes from an IBM Press Release concerning their eServer line. They would know right!
...
The latest eServer line, which will be available globally on August 31, uses from two to 16 Power 5 processors per server computer and uses as few as one-fourth the number of chips as comparable machines from rival vendors, Armonk, New York-based IBM said in a statement.
I wonder if the aug 31st date is significant?
Isn't paris expo then????
ooh goose bumps!
For a dual core version of the current chip with SMT and 1 meg L2 cache you're looking at around 125 million transistors compared to the current 58 million. Note though you're doing a bit more than just multicoring it in that example.
Adding a second core and changing nothing else would take it to around 90 million.