Power5 SMT to boost performance 40% (in threaded apps)

Discussion in 'Macintosh Computers' started by Cubeboy, Aug 20, 2003.

  1. Cubeboy macrumors regular

    Mar 25, 2003
    Bridgewater NJ

    This is considerably lower than original estimates but impressive nonetheless, especially in comparison to the Pentium 4's SMT which at most boosts performance by around 30% in well threaded programs. It does however, cast some doubts on whether the Power5 will really be 4 times faster than the Power 4 (since SMT was a big part of this claim).
  2. Lanbrown macrumors 6502a

    Mar 20, 2003
    SMT helps but is not the end-all. 2004/2005/2006 will really be interesting as new technologies get introduced to maximize the potential of processors. So far it looks like Sun will lead the pack, IBM in second and Intel pretty much near the back. Needless to say, whatever Apple will use from IBM in the future will definitely be better then what peecee users will have. Intel looks more towards clock speed then anything else; look at the power figures, ouch. If they continue the upward trend, power companies might need to put boxes on Intel machines so they can turn them off to conserve power like they do for AC units. :D

    Back on topic. A processor spends most of its times waiting for something to work on. Reducing the time it has to wait is how you can get some serious performance. This is what Sun is working on. Four times the performance is a high expectation; I expect IBM will be able to get about twice the performance from the Power5 over the Power4.
  3. Powerbook G5 macrumors 68040

    Powerbook G5

    Jun 23, 2003
    St Augustine, FL
    At any rate, all I can say about this is "Sweet."
  4. Cubeboy thread starter macrumors regular

    Mar 25, 2003
    Bridgewater NJ
    Heh, looking back at Sun's track record as well as their current standings performance-wise, you'll have to excuse me when I say, I'll believe it when I see it. Creating a roadmap and following up on it are two different things as Motorola has so painfully showed us. UltraSPARC IV was supposed to be released the first half of this year at 1.2 GHz using .13 micron process and it still hasn't been released yet (even with Intel preparing to release .09 micron cpus). I can't say I'm very impressed with their roadmap either, even UltraSPARC V (1.8-3 Ghz, .10 micron process) doesn't look that impressive considering it will be released in 2005 (against the likes of a 3+ GHz Power5+/Power6, abd 2.5+ GHz GHz Itanium)

    I wouldn't doubt however if IBM leads the pack for some time with Power5+ and Power6. Of course, Intel has, their Tanglewood processor which is being developed by the former Alpha EV8 team and is rumored to be ten times as fast as current Madison Itanium 2.

    Alpha sure thought SMT was a pretty big thing, one of the premier technologies on their EV8 processor was (the than unheard of) SMT. I would have to agree with them, the potential of SMT is enormous, right now we only have 2x hyperthreading in which it allows for 2 virtual cores to operate independently on a single cpu. Pretty soon (as in early 2004) we will begin seeing processors with 4x hyperthreading which will effectively make CPUs capable of running 4 virtual cores. It will be quite a harrowing experience to see the day when we have 8 or 16 virtual processors all operating independently on a single cpu.

    Leakage is going to be a huge problem for all chips when they go down to 90nm process. Especially during the beginning phase when the technology is still immature. SOI should partly help in this although it won't be a effective solution. Intel also isn't necessarily stuck with the Pentium 4/5, Deerfield Itanium 2s and Pentium M are alternatives. Of course, looking over some initial analysis of the Pentium 5, there is little reason to believe that Intel would implement them for anything other low power solutions.

    These days, systems seem to benefit more from lowering latency than increasing bandwidth. Just some food for thought. ;)
  5. Lanbrown macrumors 6502a

    Mar 20, 2003
    MHz doesn't matter, MHz consumes more power.

    And the Intel .09-micron processors are also going to consume 100 watts.

    You say Sun doesn't have a good track record, neither does Intel. Both of their processors were late. Sun knows what it needs to do as they have done it before. Sun will have the fastest systems by 2006, processor speed is just one factor is the overall design. In 2005, Sun will have an 8-core chip capable of 32 threads.

    So SOI is the answer? Not entirely true, Ti that makes the chips for Sun, Ti has a technology that is better then SOI. Actually IBM has one better then SOI as well, but what Ti has is even better.

    That is exactly what I said about the processor spending most of its time waiting for data. Sun knows how to lower the latency, which will be implemented in 2004. Just wait for the H series to come around, especially version 2 in 2005/2006. Version 1 should be two times faster then the Power4, but only running at 1.2GHz.
  6. Cubeboy thread starter macrumors regular

    Mar 25, 2003
    Bridgewater NJ
    As far as I know, MHz does matter, if the opposite were true, why spend the resources to produce G4s/G5s with *higher* MHz at all? The amount of power MHz consumes would actually depend on the processor. The current Pentium 4 is clocked 40% faster than the current Athlon and it sure doesn't produce 40% more heat.

    Can you list a recent Intel processor that was late?

    Exactly what cpu are you talking about? The latest SPARC on every roadmap I've seen is the US-V and the only details on it is that it will scale from 1.8-3 GHz.

    Is there a set date in 2006 by which all Sun systems will automatically become the fastest systems in the world or does it just happen when 2006 begins? :D:

    I suggest you reread my post again, more carefully this time, I never said SOI was going to be the answer, in fact I explicitly stated that it wasn't. I doubt the technology being thought up by Texas Instruments will be a solution either.

    No, you pasted a vague description, about the processor having to always wait on something. Than you try to elaborate by putting only Sun into the equation when *everyone* is trying to do the exact same thing. Tell me, do you know this for a fact? Sun seems to be focusing on lessening the time the processor has to wait on other components in the system, not the time the CPU has to wait on *itself*.
  7. Lanbrown macrumors 6502a

    Mar 20, 2003
    If you take two identical chips, except for MHz, the faster the clock rate, the more power it consumes. The G5 comes in three flavors, guess which one produces the most heat and takes the most power? Last I saw, the fp on a 1.2GHz USIII was 1118, compare that to a 3.06GHz XEON at 1003. A chip running less than half the speed is 10% faster. So MHz does matter huh?

    Yeah, the Itanic, it was late. Sun will become the fastest through overall system design, just not processor.

    You also brought up Intel going to .09. That's fine, their Itanic is still at .18 as is the Power4 from IBM. Sun is at .13 for the UltraSPARC III and IIIi.
  8. Cubeboy thread starter macrumors regular

    Mar 25, 2003
    Bridgewater NJ
    Actually, if you took the time to find the top Xeon SPECfp, instead of looking for a score that the USIII can top, you'll find it to be significantly higher than the 1.2 GHz USIII. Would you like to delve into SPECint as well?

    And yes, it would be quite obvious that MHz does matter, would you expect a 1 GHz USIII to perform the same as a 1.2 GHz model? I think your losing your grip on reality. :p

    Do you think Itanium 1 to be a recent cpu? Can you find any (and I repeat again) recent Intel cpus that were delayed?

    Again, can you elaborate on Sun's overall system design? How does it stand against RapidIO being designed by IBM and PCI Express being designed by Intel? Can you explain what "overall system design" has to do with latency which has to do with the processor?

    I think you should pay more attention to the outside world, Power4 has been at .13 with the Power4++ for quite a while and so has Itanium 2 with the Madison Itanium 2, and yes, both beat the crap out of USIII chips.
  9. Lanbrown macrumors 6502a

    Mar 20, 2003
    You can't use MHz to compare a processor from company A to processors from company B.

    You bring up the fp that doesn't totally use the cache.

    Itanium 2 was pushed back because Itanium 2 was late. Just like Sun pushed back the US IV because the III was late.

    PCI Express and RaipIO are fine, that is the bus for the peripheral. There are other times the chip must wait for data, like getting it out of memory. Or when it has information that is does not need, it takes more then one-clock cycle to go get it. Also, I said overall system design. All components play a part in overall system performance. Intel is focused on making a chip that is a strong performer, but the cost is more power consumption and heat generated. Just wait, next year will be the turn around for Sun, 2005 will bring even more and 2006 that will be back to the top.

    Last chip roadmap I saw from Intel showed the Itanic at .18. Ti is already has .09 parts available; Intel has yet to release anything.

    "While the Itanium 2 processor represents a major step in the Itanium Processor Family, work has already begun on the next generation of processors. Since the Itanium 2 processor was designed in the Intel P858 process, late in that process’s life, it is reasonable to move the Itanium 2 microarchitecture into the latest technology. That enhancement will allow generation of both lower power and higher performance
    products. That is the goal for the next-generation Madison and Deerfield products. Madison will increase performance by increasing frequency and the L3 cache size, and Deerfield will provide a lower power alternative"

    I don't see the higher cache or lower powered variant and the P858 process is .18.

    Define beat the crap out of the USIII? Lets look at the number of programmers that can code for the USIII. EPIC is unknown to most, but has been around since the 70’s. There are many more 64-bit apps available for Solaris then for the Itanic.
  10. Cubeboy thread starter macrumors regular

    Mar 25, 2003
    Bridgewater NJ
    Sure you can if both processors have similar clock to clock performance. :) Besides, if you actually read my post, you would have known that it was never my point to compare different CPUs by MHz in the first place. :D:

    Okay, no logic in this statement, how do you not "totally use the cache"? All benchmarks in the SPEC CPU2000 suite have a memory footprint of between 100 MB and 200 MB. Unless your processor is a Power4, of course it's going to take up all the cache. :confused:

    What? Itanium 2 was pushed back because Itanium 2 was late? :rolleyes: Get your facts straight, Intel has publically stated on record that Itanium 2 will be released second half of 2002 and true to word, it was released on July 8th 2002, which is well within the time frame originally planned. Sun didn't push USIV back, it wasn't released on the set timeframe or in other words delayed.

    Both PCI Express and RapidIO are positioned as general purpose interconnects meant to dramatically improve inter-system communications. This includes improving communication between the CPU and the memory, the graphics card, the hard drive, the local I/O, as well as all the peripherals.

    Tell me, do you really believe Sun is going to develop it's own line of Graphic Cards, Hard Drives, RAM, etc? What makes you so sure they'll be better than the ones produced by companies who specialize in these? I think you need to face reality.

    So far, you've still haven't provided any new information on exactly *what* Sun is working on. Nor have you provided a explanation of how you can improve latency by improving inter-system communications.

    As I've stated before, you need to read the news, Madison Itanium 2s were released in the first week of July this year as this news artile details:


    "7/3/03 — Intel this week delivered its third-generation 64-bit processor, the "Madison" release of the Itanium 2 processor. Major server vendors including Dell, HP and IBM are poised with systems built on the higher performance chip."

    SGI, HP, and IBM have had Madison servers for quite a while now. All official SPEC scores have already been posted, TPC-C scores have been posted long before the "official" release. Again, you should really pay more attention to the news before writing pointless ignorant statements like this one. :D

    Both Intel and IBM have have .09 nm parts as well as entire processors for quite a while now (in the form of prototypes, engineering samples), and soon they will have mass produced 90 nm processors and standardized them. IBM already mass produces 90 nm Xilinx Field Programmable Gate Arrays. Ti expects volume production of 90 nm communication chips (that are about as complex as a FPGA) in 4Q 2003. Intel expects volume production of it's Pentium 5 (much more complex than Ti's communication chips or IBM's FPGAs) in 4Q 2003. Who do you think is ahead?

    I'm surprised that I have to point out that "beat the crap out of USIII" refers beating the crap out of the USIII in application performance. I thought it was pretty obvious what I meant but I guess there are just some people who can't take a hint. :rolleyes:
  11. macrumors12345 macrumors 6502

    Mar 1, 2003

    I have read at least one article since the HotChips conference that said IBM reiterated the "4 times the performance of the original Power4" promise, even as they claimed "only" a 40% increase in performance from HT. Unfortunately, I can't find references for any of these articles, except for one on Google News that does clearly say that, but unfortunately the link is broken.

    I wouldn't get too excited about 4 threads per core (vs. the current 2). IBM stated that there were diminishing returns to more than 2 threads (which is why they stuck with 2), and I have heard this from other ppl too (that going with too many threads per core can actually be a bad thing, since there is too much contention for shared resources between the different threads).

    Also, keep in mind that while previous P4s and Athlons have generated almost 100 watts, the big issue with Prescott is that it is generating 103 watts *on a new process*. This is an issue in terms of further scaling; in the past P4s and Athlons have generally dissipated "only" 60 or 70 watts when moving to a new, more efficient process.


    What in the past decade would give you faith in Sun's semiconductor division is beyond me. The 8 core, 32 thread chip that you refer to is "Niagara" (I don't know what its UltraSPARC designation would be). It supposed to ship in like 2005 at the earliest, if it's on time (unlikely). Unfortunately, it's probably too little, too late, and it will almost surely suffer from the same shared resource problems that I mentioned two paragraphs above. But hey, that is just what my friend at Sun tells me, so believe what you want.

    Also, just FYI, the latest version of Itanic 2 (i.e. the 1.3-1.5 Ghz version) uses a 0.13 micron process.
  12. Lanbrown macrumors 6502a

    Mar 20, 2003
    Niagara is phase two, phase one is next year with Gemini. Both processors will be in the H series. Gemini has a dual USII care running at 1.2GHz and will be pin compatible with the IIIi. Niagara will be much faster and I assume will be based upon the USIII core.

    A lot of companies have had semiconductor problems. The Intel and AMD war did very little for the peecee market, yeah, they got high clock rate chips, but also took a lot of power, generated a lot of heat and were inefficient per clock cycle, because consumers believed that MHz is the benchmark. IBM has had their manufacturing problems as well, like cache.

    Sun is no longer trying to make the fastest processor out there, and is looking towards the whole system. Intel can't do that, only Sun, IBM and SGI can. SGI will be left out because who knows where they are headed. IBM is backing their Power series and Sun is backing the SPARC. Both are selling systems as they have their own customer base. Telco carriers like Sun equipment. EBay likes Sun, IBM and HP/Compaq. How much longer for HP/Compaq is unknown, as they were looking at HP and Sun for the front-end upgrade, IBM in the middle with the dynamic content with Sun for the backend. Why wasn't IBM in the mix for the new front-end?

    Sun has had their problems and is working on fixing them; they just need a few years. They have a huge customer base which works in their favor. If you look at the SPECfp_rate2000 scores for multiprocessor systems, how is Sun lagging behind? You can take a 32-processor system from IBM and Sun and they are close. The IBM is 10% faster, has a 500MHz faster clock cycle and two cores per processor. As the number of processors grows, Sun starts to shine. Their (Sun) processors also consume less power, thus lowering the TCO and allows more processors per area.


    Itanic was late, that pushes ALL future versions back. You can't have the successor our before the predecessor. If the roadmap showed version two coming out 18 month after the first version, and the first version is 18 months late, you don’t release both at the same time, as the successor isn't even available. Look at their old roadmap. Actually, they really don’t have a roadmap as they constantly change it, which negates the whole purpose of a roadmap.

    There are more then just RapidIO and PCI Express. I keep saying overall system performance, reducing latency is just one of those.

Share This Page