Is the current rate of PPC CPU upgrades sustainable?
If everything is to be believed, then then IBM hass been doing an incredible job of PPC development. 2.5GHz G5's were sampling in May. the 90nm G5 will be available sometime around January. The 980 is coming and will offer even more speed.
All this is great news ans a welcome departure from Motorola's endless cycle of underdelivering. IBM deserves all the kudos it gets from the Mac community.
However, I want to ask the question of how sustainable this cycle is. Consider this: the G5 is basically a stripped down Power4 with Altivec added on. The 980 will be a stripped down Power5 with Altivec also added on. Anybody heard about the Power6 lately? I think not.
Apple has had the good luck to switch to IBM CPUs right as IBM was in the process of beginning to move to the Power5 from the Power4. As such, Apple is able to rev the CPU in their high end machines rather quickly because these CPUs are derived from IBM's server chips. The G5 may live in high end desktops for only a bit over a year (it will undoubtedly live for quite some time in iMacs and portables).
So once the 980 comes out, what then? It could be quite some time until the Power6. Server CPUs require quite abit more work than desktop CPUs.
Now I don't doubt that IBM won't be able to speed up the 980 to faster speeds. I just think that the 980 will be around in desktops for a bit longer than the 970 will have been in desktop machines.
So my point is that the current rate of PPC upgrades isn't sustainable because we are currently at a rather nice time in PPC development. We shouldn't get used to IBM coming out with a new desktop chip every year because that isn't a rate that they will be sustaining. Thoughts?
All this is great news ans a welcome departure from Motorola's endless cycle of underdelivering. IBM deserves all the kudos it gets from the Mac community.
However, I want to ask the question of how sustainable this cycle is. Consider this: the G5 is basically a stripped down Power4 with Altivec added on. The 980 will be a stripped down Power5 with Altivec also added on. Anybody heard about the Power6 lately? I think not.
Apple has had the good luck to switch to IBM CPUs right as IBM was in the process of beginning to move to the Power5 from the Power4. As such, Apple is able to rev the CPU in their high end machines rather quickly because these CPUs are derived from IBM's server chips. The G5 may live in high end desktops for only a bit over a year (it will undoubtedly live for quite some time in iMacs and portables).
So once the 980 comes out, what then? It could be quite some time until the Power6. Server CPUs require quite abit more work than desktop CPUs.
Now I don't doubt that IBM won't be able to speed up the 980 to faster speeds. I just think that the 980 will be around in desktops for a bit longer than the 970 will have been in desktop machines.
So my point is that the current rate of PPC upgrades isn't sustainable because we are currently at a rather nice time in PPC development. We shouldn't get used to IBM coming out with a new desktop chip every year because that isn't a rate that they will be sustaining. Thoughts?
Comments
By going 90 nm power consumption will go down a bit. But Moore's law says: double number of transistors every 18 months. We're at, what, 60 million now? Every transistor is (put simply) a little bit of heat. Making them smaller reduces their heat output. However, increasing their number increases the total heat output. So it won't be long before 90 nm just doesn't sweat it, and production stalls because the procs start melting everything inside a tower, or the towers start moving around as hovercrafts because of the industrial strength ventilation they require. As the situation is now, I believe the next target is a 65 nm process.
This isn't endless. New concepts shall and will be necessary if the race wants to continue. I wonder which, personally. Stuff like multi-processor systems are fairly obvious. What else? Maybe time and some other knowledgeable member will tell.
About the feet: (simply put) every foot of a processor requires a certain amount of power to be pushed into the processor's socket. More transistors -> more feet -> more required power. You can only push so hard on that processor however before you plain crush it.
Challenges...
Originally posted by Yevgeny
If everything is to be believed, then then IBM hass been doing an incredible job of PPC development. 2.5GHz G5's were sampling in May. the 90nm G5 will be available sometime around January. The 980 is coming and will offer even more speed.
All this is great news ans a welcome departure from Motorola's endless cycle of underdelivering. IBM deserves all the kudos it gets from the Mac community.
However, I want to ask the question of how sustainable this cycle is. Consider this: the G5 is basically a stripped down Power4 with Altivec added on. The 980 will be a stripped down Power5 with Altivec also added on. Anybody heard about the Power6 lately? I think not.
Actually, IBM has mentioned it. It's obviously still on paper, but they're working on it. IBM is very keen on letting people know that they're shifting into overdrive.
Also, there's Cell. If I'm not mistaken, the earliest fruits of that project are sitting at #73 on the supercomputing list, with consoles and other applications (and #73's big brother supercomputer) to follow.
So once the 980 comes out, what then? It could be quite some time until the Power6. Server CPUs require quite abit more work than desktop CPUs.
On the other hand, you can save yourself a lot of work by doing things right the first time. How many revisions has AltiVec been through? POWER5 is based on POWER4; POWER6 will probably inherit a lot of POWER5. A lot of this is possible because the POWER (and PowerPC) specs were done right the first time, and because IBM did the implementation right. They don't have to start each iteration from scratch. I believe POWER6 is scheduled to appear in 2006, and Fishkill is good past 65nm, so we'll see a lot of work done. Also - and this is crucial for Apple - IBM is geared up to be able to rapidly develop custom CPUs with high bandwidth interconnects and SMP capabilities. They don't have to make gigantic, centralised jacks of all trades like the Itanic.
Now I don't doubt that IBM won't be able to speed up the 980 to faster speeds. I just think that the 980 will be around in desktops for a bit longer than the 970 will have been in desktop machines.
So my point is that the current rate of PPC upgrades isn't sustainable because we are currently at a rather nice time in PPC development. We shouldn't get used to IBM coming out with a new desktop chip every year because that isn't a rate that they will be sustaining. Thoughts?
It's true that there won't be a new POWER line every year. They're clearly trying to get 2-3 years out of each generation (they have to, because that's about how long it takes to design the next one!). But if you look at what IBM can do with, say, the 440 core, they can get all kinds of use out of one design within 2-3 years. It's not unreasonable at all to expect that IBM will keep things coming at a pretty good pace. You can go dual core or MCM, enlarge caches, tweak the design, shrink the process, put the memory controller on die, etc.
Here my thoughts for 2004 :
Well let's say a P4 3 ghz equal a G5 2 ghz
-A G5 3 ghz will be rughly 50 % more performant than the 2 ghz (approximatively, it will depends of the bus, memory, and the L2 cache)
- a P4 prescott is ten % more performant than a P4 northwood at equal mhz. So a P4 prescott at 4 ghz is approximatively 50 % more performant.
- dual xeon, are less performants than dual G5 in MP aware software.
So we can assume that the 3 ghz G5 will have the same level of performance than a 4 ghz Prescott.
Now a PPC 980 will be 25 % faster (if we take for reference the power 5 vs the power4) than a ppc 970.
So the PPC 980 will be equal to a 5 ghz Prescott. Considering that the PPC 980 have far more chances to ship before a prescott 5 ghz, and considering that the dual works more efficiently on the mac platform, than in the Intel one, we can assume that IBM will have the edge over intel for the next coming year.
We don't know the future for 2005 or 2006. But i will say that the G6 will not come before 2006. It will be based upon the power 6. The only rumors we get, was that the power6 will feature a new SIMD unit with VMX 2 and many more instructions.
There is also a possibility that the PPC 980 have two versions : the classical one pin compatible with the ppc 970 and a new one with built in memory controller (that will requiere a brand new mobo). As I consider that the current generation of mobo will last from 18 to 24 months, this new version will not arrive before 2005. As Amorph stated there is also the place for a PPC 9800 : a dual core variant of the PPC 980 will kick the ass of any P4 chip. Why not a PPC 9800 on 65 nm process ?
And IBM has already mentioned the Power6 http://zdnet.com.com/2100-1103-892836.html . It looks like they have maintained a fast pace since the Power3.
Maybe you should do a little research and get some facts, before critizing. I don't see any indication that this was an abnormally "nice time in PPC development". This looks like a normal evolutionary cycle to me.
Originally posted by Leonard
If you ask me, IBM can sustain the current rapid pace of PowerPC development. When did the Power3 first ship? Something like 1998? The Power4 must have been around in 2000, 2001. The Power5 is coming in 2003, 2004.
Ok, so 2004 minus 1998 is 7 years (don't forget Y2k). 7 years from debut of the Power3 to the Power5, 3 CPUs. That means a bit over 2 years between CPU cycles (7/3 = 2.33). Now Apple will go from the 970 to the 980 in one year which is half the time that IBM took to develop the next generation Power CPU. So, to use your own facts to disagree with you, no IBM can not output a new 9X0 chip every year because it takes them two years on average to make the server version of the chip that they turn into the desktop version. Thanks for the stats.
Originally posted by Leonard
Maybe you should do a little research and get some facts, before critizing. I don't see any indication that this was an abnormally "nice time in PPC development". This looks like a normal evolutionary cycle to me.
I am not criticising. I am asking if the current rate of PPC development is sustainable, or in end user terms if we will have a G7 CPU (PPC 990) announced in the summer of 2005 (because that is what we seem to be on track for at the current rate).
I think that we will have many revisions of the 980 and that the 980 will be the speed king for more time than the 970 will be the speed king. I think that IBM is more than capable of keeping the chips fast, but that fundamental revisions of the core will be less frequent. Mac users may have to satisfy themselves with speed boosts instead of brand new CPUs. It's a tough life
Oh, and as a minor semantic peeve, microprocessors do not "evolve". They are designed. You wouldn't want a CPU that was trying to evolve into a better CPU by adding some random transistors here and there. I can't stand it when people talk about software and hardware evolution. If this stuff is evolving, then how come I am stuck debugging it and writing it?
Originally posted by Yevgeny
Now Apple will go from the 970 to the 980 in one year which is half the time that IBM took to develop the next generation Power CPU.
You're working on the assumption that an actual successor for the 970 will be coming next year and I wouldn't be betting on that. An improved version of the current chip yes but not a successor. In fact I'd be betting quite heavily that the successor will appear in the 2nd half of 2005 and be a dual core chip but hey that's me for you
Originally posted by Yevgeny
Ok, so 2004 minus 1998 is 7 years (don't forget Y2k). 7 years from debut of the Power3 to the Power5, 3 CPUs. That means a bit over 2 years between CPU cycles (7/3 = 2.33). Now Apple will go from the 970 to the 980 in one year which is half the time that IBM took to develop the next generation Power CPU. So, to use your own facts to disagree with you, no IBM can not output a new 9X0 chip every year because it takes them two years on average to make the server version of the chip that they turn into the desktop version. Thanks for the stats.
Actually, the way I see it is this. For years, we were under the thumb of Motorola's incompetence. The way Motorola had us, we as a community was doomed to go down. Apple did the very best it could to keep itself afloat, with rather poor scaling chips.
When IBM and Apple joined forces again, it really was a godsend. I honestly think that Apple may in fact release a 980 maybe as soon as 2004 (more than likely 2005). The reason why is to get their machines away from Motorola's chipset. After that, we won't be complaining about the chips, because they will be going faster due to scaling, not necessarily chip design.
Oh, and as a minor semantic peeve, microprocessors do not "evolve". They are designed.
Oh, the minor things we pick upon... As a programmer, any kind of process in which we better it, we call an evolution. To take it directly from the American Heritage Dictionary, Fourth Edition...
Evolution - n.
A gradual process in which something changes into a different and usually more complex or better form.
Base word: Evolve
By this definition, I would say that the process from G4, to 970, to 980 would definatily classify itself as an "evolution".
Originally posted by Yevgeny
Now Apple will go from the 970 to the 980 in one year which is half the time that IBM took to develop the next generation Power CPU. So, to use your own facts to disagree with you, no IBM can not output a new 9X0 chip every year because it takes them two years on average to make the server version of the chip that they turn into the desktop version. Thanks for the stats.
Your welcome for the stats, but ss someone else pointed out, your making an assumption the 980 is coming out in 2004. It might not.
I think that we will have many revisions of the 980 and that the 980 will be the speed king for more time than the 970 will be the speed king. I think that IBM is more than capable of keeping the chips fast, but that fundamental revisions of the core will be less frequent. Mac users may have to satisfy themselves with speed boosts instead of brand new CPUs. It's a tough life
Why do we need chip revisions so often? There's nothing wrong with just healthy speed increases. IBM has lots of chip technologies that can eek out that extra bit of speed. Who knows what the maximum Ghz is of the 970.
Oh, and as a minor semantic peeve, microprocessors do not "evolve". They are designed. You wouldn't want a CPU that was trying to evolve into a better CPU by adding some random transistors here and there. I can't stand it when people talk about software and hardware evolution. If this stuff is evolving, then how come I am stuck debugging it and writing it?
Well, I guess us computer folks believe CPUs evolve as their design change and improve them. Humans and other animals don't necessarily evolve themselves.
The 970 is in many ways a superset of the Power4 and even has the built in ability to support an 8-way configuration and actually allows for sharing both the L2 and L1 caches between the SMP processors. The G5 actually does this on a dual setup. So even if you are only running a single-threaded app, that App has twice the L1 and L2 cache available to it on a dual G5. Pretty cool.
The 970 90NM SOI implementation goes even a step further and adds a new power management technology called PowerTune, which we currently have little information on, but apparently has been developed specifically for the mobile market, something the Power4 will never find itself in. The 970 90NM should also be faster due to shorter wire lengths and other refinements in addition to a higher clock. And I dare say that the 90NM processor will be cheaper as IBM will be able to significantly increase their yields per wafer, so we will probably see it in iMac's and am hoping a very low power version in iBooks, so Apple can be purely on 64-bit pocessors and phase out the 32-bit G4 support within 5 years of that event. It will make for a cleaner framework in 2009.
The 980 was developed simultaneously with the Power5 and while the SMT, doubled L1 and L2 caches and built in memory controller information is out, it is also rumored to have the next generation of Velocity compatible SIMD, more than doubling its SIMD capability. It should be an awesome processor and out by the the end of Summer at the latest. Speed will not be 25% faster, but at least 100% faster than the current 970. The Power5 is up to 400% faster than the Power4 in way of reference.
Originally posted by stingerman
Actually there were three major revisions of the Power3 before the Power4 came out and there is actually a Power4+. The 970 is not a subset of the Power4, it is built up from the Power4 core, but additional instructions and functionality was added to the 970, and not just Altivec.
Hmm... according to Henriok's chart, the Power3 only had two revisions, a 250 nm version and a 200 nm version. The Power4 is a 180 nm processor while the Power4+ is a 130 nm processors. So, all IBM is doing is just fabbing the architecture on more advanced nodes. So, there is a high probably that a 90 nm 970 will be called a 970+.
The 970 is in many ways a superset of the Power4 and even has the built in ability to support an 8-way configuration and actually allows for sharing both the L2 and L1 caches between the SMP processors. The G5 actually does this on a dual setup. So even if you are only running a single-threaded app, that App has twice the L1 and L2 cache available to it on a dual G5. Pretty cool.
I really doubt they share the on-die cache in the manner you are talking about in the Power Mac G5. There really is no point to doing it that way. Shared cache coherency is maybe what you are talking about though.
The 970 90NM SOI implementation goes even a step further and adds a new power management technology called PowerTune, which we currently have little information on, but apparently has been developed specifically for the mobile market, something the Power4 will never find itself in.
The Power5 will have good power management capability. So, these power management features are going to be in the 90nm 970? It would be good news if so. It would be good news if they have a 130 nm version too.
Speed will not be 25% faster, but at least 100% faster than the current 970. The Power5 is up to 400% faster than the Power4 in way of reference.
At the same clock rate, multithreading will supposedly give the Power5 40% more performance than the Power4 on multithreaded applications. Definitely need the power management too, since multithreading will require more power.
I really doubt they share the on-die cache in the manner you are talking about in the Power Mac G5. There really is no point to doing it that way. Shared cache coherency is maybe what you are talking about though.
There is nothing to doubt, it is what it does. Read Apple's own G5 Architecture documentation. Basically a dual G5 behaves similarly to a dual core Power4 both combining their caches. The Interconnects exhibit the qualities of a bus with the advantages of switching. It is really an advanced design that benefits an App that can use the additional cache. By the way, the Opteron does the same trick.
Originally posted by THT
Hmm... according to Henriok's chart, the Power3 only had two revisions, a 250 nm version and a 200 nm version. The Power4 is a 180 nm processor while the Power4+ is a 130 nm processors. So, all IBM is doing is just fabbing the architecture on more advanced nodes. So, there is a high probably that a 90 nm 970 will be called a 970+.
There were 3:
1. Power3 - I
64-bit addressability, double-word interger operations, and symmetric multiprocessor support in the PowerPC Architecture. The technology strategy of the POWER3 design was to produce a highly sophisticated processor core and memory subsystem in an advanced, but well-established technology.
2. Power3 - II
POWER3-II design is the next step, planned to result in an increase of frequency by up to 50% by tuning the design and moving into IBM's cutting-edge copper technology?CMOS7S.
3. Power3 - III
The POWER3-III design is step three, with plans for increased frequency by as much as 40% over the POWER3-II architecture, with more design tuning combined with a move to IBM's newest breakthrough technology - Silicon on Insulator (SOI).
http://www-1.ibm.com/servers/eserver.../power3wp.html
Originally posted by THT
The Power5 will have good power management capability. So, these power management features are going to be in the 90nm 970? It would be good news if so. It would be good news if they have a 130 nm version too.
The information on PowerTune came out not too long ago. I originally posted it at ThinkSecret.com in their hardware forum under the 970 90NM SOI, the references to back it up are there.
Here is the link if you want to get the supporting information:
http://www.pstheme.com/cgi-bin/think...ST;f=14;t=1764
But from a non-tech perspective, it looks like IBM has every incentive to drive the PPC line at full steam for the next few years at least. Look at this article/rumor from yesterday on the Inquirer:
http://www.theinquirer.net/?article=12693
We already know about IBM's blade servers, but if IBM really is switching over to Linux on all of its desktops, could they be looking to have PPC boxes running Linux?
It seems so long ago and the details are hazy to me now, but I remember the first signs of trouble for Apple and the PPC line was reading about Motorola's management dumping all their Macs and replacing them with x86 windows boxes. It was clear from that point on that Motorola had almost no interest internally in driving their desktop chip line forward.
x86 land looks to be a complete disaster right now. People are wondering why the Itanium line isn't just canceled by Intel. Intel has been trying to downplay 64bit desktop chips, but now looks like they are scrambling to tack on 64bit extensions in response to AMD's latest chips. It sounds like Intel is having major heat and process problems.
It looks like we can sit back enjoy the ride for the next couple of years with a steady flow of significant chip/system upgrades every six months from IBM and Apple.
Originally posted by stingerman
3. Power3 - III
The POWER3-III design is step three, with plans for increased frequency by as much as 40% over the POWER3-II architecture, with more design tuning combined with a move to IBM's newest breakthrough technology - Silicon on Insulator (SOI).
To this day, I have not found any evidence that IBM ever shipped a 180 nm Power3-III. They were slated to be in IBM's Nighthawk 2 servers, but it appears the Power4 came out in time and IBM decided to stick with the Power3-II for their entry servers.
If you can find evidence of the Power3-III in an actual product, let me know.
Originally posted by stingerman
There is nothing to doubt, it is what it does. Read Apple's own G5 Architecture documentation. Basically a dual G5 behaves similarly to a dual core Power4 both combining their caches. The Interconnects exhibit the qualities of a bus with the advantages of switching. It is really an advanced design that benefits an App that can use the additional cache. By the way, the Opteron does the same trick.
There is everything to doubt, since it doesn't make any sense to do it that way, share the L1 and L2 cache in SMP configurations so that one processor can use the cache of another processor. If you know which link it is, I'd like to see it.
And I would doubt the Opteron does it the same way as well.
Originally posted by Tuttle
Chip manufacturing and process tech isn't my area, so I don't have anything technical to add.
But from a non-tech perspective, it looks like IBM has every incentive to drive the PPC line at full steam for the next few years at least. Look at this article/rumor from yesterday on the Inquirer:
http://www.theinquirer.net/?article=12693
We already know about IBM's blade servers, but if IBM really is switching over to Linux on all of its desktops, could they be looking to have PPC boxes running Linux?
It seems so long ago and the details are hazy to me now, but I remember the first signs of trouble for Apple and the PPC line was reading about Motorola's management dumping all their Macs and replacing them with x86 windows boxes. It was clear from that point on that Motorola had almost no interest internally in driving their desktop chip line forward.
x86 land looks to be a complete disaster right now. People are wondering why the Itanium line isn't just canceled by Intel. Intel has been trying to downplay 64bit desktop chips, but now looks like they are scrambling to tack on 64bit extensions in response to AMD's latest chips. It sounds like Intel is having major heat and process problems.
It looks like we can sit back enjoy the ride for the next couple of years with a steady flow of significant chip/system upgrades every six months from IBM and Apple.
You make an interesting point regarding IBM's transition to Linux. Since IBM is slowly starting to ditch Microsoft (at least internally), it does indeed mean that they could potentially use a hardware PPC platform. As well, if IBM is doing it internally, I see no reason why IBM wouldn't sell such systems externally.
In fact, it would be worth checking through the IBM product listings. PPCs are obviously used in their server products, but I wouldn't be suprised if there is already a desktop product available.
Even more interesting would be Apple OS X on IBM PPC platform, though I have some pretty strong reservations about the chances of this happening (again). This was already looked at previously back in the Apple clone days, the platform was called CHRCP or something.
Originally posted by Chagi
Even more interesting would be Apple OS X on IBM PPC platform, though I have some pretty strong reservations about the chances of this happening (again). This was already looked at previously back in the Apple clone days, the platform was called CHRCP or something.
I think you may have one too many "C"s in there. I think it was called CHRP - Common Hardware Reference Platform, if memory serves me. Most people pronounced it as "chirp".
Originally posted by THT
There is everything to doubt, since it doesn't make any sense to do it that way, share the L1 and L2 cache in SMP configurations so that one processor can use the cache of another processor. If you know which link it is, I'd like to see it.
And I would doubt the Opteron does it the same way as well.
Dude, you got to do your own due diligence. Read the white papers on the Opteron and the G5, respectively on the AMD and Apple sites. Your thinking old processor interconnect technology, the G5 and the Opteron use their own proprietary though similar coherent processor interconnect technology that is derived from the hypertransport non-coherent interconnect spec (hypertransport doesn't currently define coherent PI's so that is left up to the designers, though both Apple and AMD are peers on the hypertransport committee and probably work closely together, thus the similarities.)
There is nothing to doubt, its a fact and part of the design. Enough said.