Multithreading and the PowerPC
Courtesy of this Ars thread.
If you're interested in performance benefits you'll probably want to check over here.
Now the question is does 10.3.5 make the distinction as well or do we have no multithreading until after Tiger is released?
Quote:
Originally posted by Catfish_Man42:
Doing sysctl -a in 10.4 turns up two interesting values, specifically the number of physical and logical processors. Since this distinction isn't made in 10.3 (as far as I can tell, anyway) does this imply SMT support in 10.4?
Originally posted by Catfish_Man42:
Doing sysctl -a in 10.4 turns up two interesting values, specifically the number of physical and logical processors. Since this distinction isn't made in 10.3 (as far as I can tell, anyway) does this imply SMT support in 10.4?
If you're interested in performance benefits you'll probably want to check over here.
Now the question is does 10.3.5 make the distinction as well or do we have no multithreading until after Tiger is released?
Comments
Originally posted by Telomar
Now the question is does 10.3.5 make the distinction as well or do we have no multithreading until after Tiger is released?
This is fallacious logic. Apple could release a modified MacOS X when POWER5-derived PPCs ship. Apple would be stupid to tie hardware to a major OS release since both have fickle attributes.
All this tidbit shows that SMT-enabled PowerPC processors are imminent. But we already knew that.
Did IBM throttle back the Apple roadmap once other larger customers jumped onboard? This apparent lack of certainty is a bit concerning.
Any further information, distant echoes, or blips on the potential radar screen? Or will IBM simply push the 970fx into the next revision?
All this tidbit shows that SMT-enabled PowerPC processors are imminent. But we already knew that.
Well I think many of us speculated that SMT was an option but here is much more proof than mere speculation. Apple has added support in the kernel for logical processors and that's quite encouraging. This is a bird in the hand and it supports the case for the 970fx being replaced by a 97x CPU. I don't suppose IBM would go through the effort of adding SMT to the 970fx being that it has but 500mhz of scaling max to go.
Hmmm while 2004 isn't looking like the hot year 2005 in my estimation is looking pretty damn good. I don't believe the rumors about Quad Processors but I do believe that Apple is testing Dual SMT enabled processors and people are simply calling them Quads. If Apple releases these "Quasi Quads" along with Tiger and some more suprises(new apps) I can definitely see Apple making some waves in 2005.
Originally posted by hmurchison
Well I think many of us speculated that SMT was an option but here is much more proof than mere speculation. Apple has added support in the kernel for logical processors and that's quite encouraging. This is a bird in the hand and it supports the case for the 970fx being replaced by a 97x CPU. I don't suppose IBM would go through the effort of adding SMT to the 970fx being that it has but 500mhz of scaling max to go.
Hmmm while 2004 isn't looking like the hot year 2005 in my estimation is looking pretty damn good. I don't believe the rumors about Quad Processors but I do believe that Apple is testing Dual SMT enabled processors and people are simply calling them Quads. If Apple releases these "Quasi Quads" along with Tiger and some more suprises(new apps) I can definitely see Apple making some waves in 2005.
I agree with you. Dual SMT processors are on the way. Scaling up CPU frequencies become harder and harder : getting a 30 % bonus of performance is something of value.
The die size of the PPC 970 FX, is small, even with SMT it will still smaller than a Prescott (and I think that there is also room for a bigger L2 cache). Most of the R&D has been done, and as many of our experts have said, IBM can develop this feature on different chips simultaneously.
I am ready to bet several bottle of Champagne that the next PPC chip will be SMT.
Originally posted by Powerdoc
I agree with you. Dual SMT processors are on the way. Scaling up CPU frequencies become harder and harder : getting a 30 % bonus of performance is something of value.
The die size of the PPC 970 FX, is small, even with SMT it will still smaller than a Prescott (and I think that there is also room for a bigger L2 cache). Most of the R&D has been done, and as many of our experts have said, IBM can develop this feature on different chips simultaneously.
I am ready to bet several bottle of Champagne that the next PPC chip will be SMT.
Benchmarks for the Power5 in a 130nm process. This thing is the fastest processor on the planet by a considerable margin:
SPEC CPU2000
__________________ L1 I/D L2/L3______ Int_________ FP
Model___#CPUs MHz___(KB)___ (MB)___ peak___base___peak___base
*p5-520 P5/1 1650 64/32___1.9/36___ --___ -- 2,138 2,034
*p5-550 P5/1 1650 64/32___1.9/36___ --___ -- 2,221 2,121
*p5-570 P5/1 1900 64/32___1.9/36___ --___ -- 2,702 2,576
SPECrate CPU2000
__________________ L1 I/D___ L2/L3_______________Int_________ FP
Model___#CPUs___MHz___(KB)______(MB) rperf___peak___base___peak___base
*p5-520 P5/2___1650 64/32___ 1.9/36___9.86___ --___ --___43.0___41.5
*p5-550 P5/4___1650 64/32___ 3.8/72 19.66___ --___ --___84.8___82.1
*p5-570 P5/4___1900 64/32___ 3.8/72 22.26___76.3___74.4 130.0 125.0
*p5-570 P5/8___1900 64/32___7.6/144 42.14 147.0 141.0 241.0 249.0
*p5-570 P5/12 1900 64/32 11.4/216 60.50___ --___ --___ --___ --
*p5-570 P5/16 1900 64/32 15.2/288 77.45 294.0 273.0 460.0 438.0
STREAM
____________________________________ Standard Tuned
_____________________L1 I/D L2/L3___ Triad___ Triad
Model___# CPUs MHz (KB)___ (MB)___ MB/sec___MB/sec
*p5-520 P5/2___ 1650 64/32___1.9/36___4,275___ 4,510
*p5-550 P5/4___ 1650 64/32___3.8/72___8,201___ 8,986
*p5-570 P5/4___ 1900 64/32___3.8/72___19,250___26,214
*p5-570 P5/8___ 1900 64/32___7.6/144 35,934___43,037
*p5-570 P5/16___1900 64/32___15.2/288 44,241___45,187
More impressive is that IBM's new 16-way machine is beating HP's current 64-way Itanium based machine. So here finally is an indication of where the Mac should be next year.
Originally posted by JRG
Benchmarks for the Power5 in a 130nm process. This thing is the fastest processor on the planet by a considerable margin
Looks like IBM decided not to release SPECint2000 scores. Thus it seems the Pentium 4 is still king of integer ops so calling the Power5 the fastest processor on the planet is misleading at best.
See here for a more detailed discussion on those scores.
AFAIK _one_ Power5 CPU have two cores doing two threads each, resulting in 4 logical processors. A dual Power5 system would consist of 8 logical processors. A 16 way Power5 system would consist of 16 CPUs, 32 actual cores and 64 logical processors.
Am I right?
The really cool Power5 MCM have 4 CPUs. Are there MCMs with just 2 CPUs on them?
Originally posted by Henriok
I'm a bit confused here.. What is considered _one_ Power5 CPU, and IBM's recent release of dual Power5 systems?
AFAIK _one_ Power5 CPU have two cores doing two threads each, resulting in 4 logical processors. A dual Power5 system would consist of 8 logical processors. A 16 way Power5 system would consist of 16 CPUs, 32 actual cores and 64 logical processors.
Am I right?
The really cool Power5 MCM have 4 CPUs. Are there MCMs with just 2 CPUs on them?
Yes on both counts.
The p5-520, p5-550, and p5-570 systems don't use MCMs; they use DCMs which each have one POWER5 chip and one L3 cache chip.