Hot Rod CPU
- Thread starter Bozo
- Start date
jtr1962
Storage? I am Storage!
You have to love marketing:
All that means is you'll have less power per square mm of chip. If you have more area, you can still have more power than today's CPUs. Not that I mind, though. All these high-powered processors mean an assortment of nice, cheap, efficient heat sinks suited for my thermoelectric module hobby. Just ordered a nice P4 copper heat sink for $17 shipped which probably would have cost $100 a few years ago, if you could even find one.
Gilbo
Storage is cool
There's a much more indepth look into the albeit limited information on IBM's Power6 at Real World Tech. Those guys really know their processors.
I've found them and ArsTechnica to be the best for indepth CPU information now that Ace's is relatively inactive. Real World Tech is more technical and thorough than anyone else, but it can be pretty dense reading sometimes.
I've found them and ArsTechnica to be the best for indepth CPU information now that Ace's is relatively inactive. Real World Tech is more technical and thorough than anyone else, but it can be pretty dense reading sometimes.
Gilbo
Storage is cool
I'd just like to add that I think it is very interesting that IBM is going in exactly the opposite direction of everyone else in the CPU business right now. It makes one wonder about the amount of thought that went into this reactionary about-face at Intel, and the "multi-threaded performance is the only way forward / single-threaded performance is dead" PR bombardment everyone's putting out.
Are the process difficulties really are so insurmountable as they've been hyped to be?
Are the process difficulties really are so insurmountable as they've been hyped to be?
Fushigi
Storage Is My Life
I wonder if it has more to do with the economies of scale. i/p/zSeries machines are in another realm entirely compared to Wintel. Things like reliability and scalability are far more important than commoditization.
Counter that to our Wintel philosophy: Buy it with the 3 year warranty and junk it when that expires. Scale out vs. up. Count on replacing, not upgrading, the OS every 3-4 years.
Also, system throughput is another hugely important factor. Some reports people run against our SQL Server based databases take 6+ hours to complete. Similar reports on the iSeries take under 20 minutes. The SQL Server will have a half dozen concurrent users; the iSeries about 250. The SQL Server will track a single client's data; the iSeries about 90. To provide perspective, our iSeries has 2 POWER5 cores active; the largest iSeries has 32 cores.
IMO IBM thinks in terms of systems and solutions, not parts. The likes of Intel and Microsoft can never do this as long as they are in the multiple supplier, commodity-price driven market. IBM creates their technology to fit a vision of the completed system. Intel, etc. creates their technology to fit a certain market (price point, formfactor, etc.).
BTW, when IBM took the AS/400 to PPC, they added a microcode routine that, upon first access, automatically converted apps from 48 bit to 64 bit. The 64 bit object was then stored and used for all successive accesses. There was no user or sysadmin intervention required; just a brief delay while the object was converted. I don't see why they couldn't do the same thing to enable porting of zSeries code to PPC.
Counter that to our Wintel philosophy: Buy it with the 3 year warranty and junk it when that expires. Scale out vs. up. Count on replacing, not upgrading, the OS every 3-4 years.
Also, system throughput is another hugely important factor. Some reports people run against our SQL Server based databases take 6+ hours to complete. Similar reports on the iSeries take under 20 minutes. The SQL Server will have a half dozen concurrent users; the iSeries about 250. The SQL Server will track a single client's data; the iSeries about 90. To provide perspective, our iSeries has 2 POWER5 cores active; the largest iSeries has 32 cores.
IMO IBM thinks in terms of systems and solutions, not parts. The likes of Intel and Microsoft can never do this as long as they are in the multiple supplier, commodity-price driven market. IBM creates their technology to fit a vision of the completed system. Intel, etc. creates their technology to fit a certain market (price point, formfactor, etc.).
BTW, when IBM took the AS/400 to PPC, they added a microcode routine that, upon first access, automatically converted apps from 48 bit to 64 bit. The 64 bit object was then stored and used for all successive accesses. There was no user or sysadmin intervention required; just a brief delay while the object was converted. I don't see why they couldn't do the same thing to enable porting of zSeries code to PPC.
Sol
Storage is cool
I think it also comes down to the specific purpose these chips are used for. The Z series have always shone in tasks like accessing hard drives which is one of the features which makes them so blazingly fast at database access.
Nobody ever tries to use a Z series for complex calculations though, for that sort of application IBM uses much the same approach as everyone else is taking on. Throw a huge number of processor at the problem... Actually IBM started down that path earlier than just about any other major player and with Blue Gene L have applied the technique on a larger scale than anyone else.
The Z series runs exactly one OS, it pumps through a huge ammount of data and it doesn't need to do a whole lot with any of that data processing wise. When you know exactly what your processor is going to be expected to do it probably makes for much simpler prediction algorythms and a much simpler logical unit. That in turn should make higher clock speeds possible with relatively lower power consumption (Since you can probably cut out a lot of transistors).
Nobody ever tries to use a Z series for complex calculations though, for that sort of application IBM uses much the same approach as everyone else is taking on. Throw a huge number of processor at the problem... Actually IBM started down that path earlier than just about any other major player and with Blue Gene L have applied the technique on a larger scale than anyone else.
The Z series runs exactly one OS, it pumps through a huge ammount of data and it doesn't need to do a whole lot with any of that data processing wise. When you know exactly what your processor is going to be expected to do it probably makes for much simpler prediction algorythms and a much simpler logical unit. That in turn should make higher clock speeds possible with relatively lower power consumption (Since you can probably cut out a lot of transistors).
Gilbo
Storage is cool
I'm not so sure myself Sol. Firstly, the Power6 is also being used in the iSeries and pSeries. Both of which are geared towards more general-purpose, server computing --particularly the pSeries.
In keeping with this, from the sounds of it, the Power6 isn't quite as wide as the Power4 & 5 --possibly slightly longer as well-- but it's still very much a general-purpose chip by all appearances. They haven't cut many transistors out of it (according to the limited but by-all-appearances accurate information available). They haven't done away with out-of-order execution; it's not an extreme adaptation like the Cell or Niagara.
Secondly, it appears that the power consumption and clock-speed achievements are being brought about primarily by process technology and custom circuit tuning of critical pathways --not so much by virtue of adaptation to a specific task. (The presence of OoO execution hardware, which is very often a limiting pathway with respect to clockrate scaling strongly suggests this, since removing the OoO hardware is one of the easiest ways to make a chip clock higher.) Process technology is something IBM has invested in tremendously and is known to be quite good with, while hand-tuning critical pathways is something that IBM has always avoided in the engineering of their processors because the effort cannot be amortized as easily over low-volume production (AMD and Intel have long done this). It's interesting that they're finally putting the effort in, and I wonder what it implies about commoditization of Power? Maybe moving the zSeries to power is just the beginning of trying to push it elsewhere as well?
In keeping with this, from the sounds of it, the Power6 isn't quite as wide as the Power4 & 5 --possibly slightly longer as well-- but it's still very much a general-purpose chip by all appearances. They haven't cut many transistors out of it (according to the limited but by-all-appearances accurate information available). They haven't done away with out-of-order execution; it's not an extreme adaptation like the Cell or Niagara.
Secondly, it appears that the power consumption and clock-speed achievements are being brought about primarily by process technology and custom circuit tuning of critical pathways --not so much by virtue of adaptation to a specific task. (The presence of OoO execution hardware, which is very often a limiting pathway with respect to clockrate scaling strongly suggests this, since removing the OoO hardware is one of the easiest ways to make a chip clock higher.) Process technology is something IBM has invested in tremendously and is known to be quite good with, while hand-tuning critical pathways is something that IBM has always avoided in the engineering of their processors because the effort cannot be amortized as easily over low-volume production (AMD and Intel have long done this). It's interesting that they're finally putting the effort in, and I wonder what it implies about commoditization of Power? Maybe moving the zSeries to power is just the beginning of trying to push it elsewhere as well?
Sol
Storage is cool
The inclusion of Power chips in the X-Box360 and the use of Cell in the PS3 have probably done more for the commoditization of the Power architechture than anything previous.
IBM is definately gearing up to push Power into any market segments it thinks possible so I guess at the moment just producing the fastest most efficient PPC chip possible is a pretty massive priority.
IBM is definately gearing up to push Power into any market segments it thinks possible so I guess at the moment just producing the fastest most efficient PPC chip possible is a pretty massive priority.