Low Latency L2 Cache

We mentioned at the start of this review that the Pentium M featured a very large, yet low latency L2 cache - at 2MB for the current 90nm Dothan based Pentium M. But how much low latency are we talking? To find out, we turned to ScienceMark 2.0 and Cachemem.

First, let's have a look at the latency of the Pentium M's 64KB L1 cache:

   Cachemem L1 Latency  ScienceMark L1 Latency
AMD Athlon 64 3 cycles 3 cycles
Intel Pentium 4 (Northwood) 1 cycle 2 cycles
Intel Pentium 4E (Prescott) 4 cycles 4 cycles
Intel Pentium M 3 cycles 3 cycles

With such a large L1 cache, it is difficult to get much lower than 3 cycles, as we see that the Pentium M has a similar L1 access latency as the Athlon 64. What is also important to note, however, is that the Pentium M does have a lower L1 access latency than the Pentium 4E.

But what we came here to look at was L2 cache latency, which matters much more in real world application performance where not everything fits into L1:

   Cachemem L2 Latency  ScienceMark L2 Latency
AMD Athlon 64 17 cycles 18 cycles
Intel Pentium 4 (Northwood) 16 cycles 16 cycles
Intel Pentium 4E (Prescott) 23 cycles 23 cycles
Intel Pentium M 10 cycles 10 cycles

Here's where things get very interesting - the Pentium M has the lowest L2 cache access time of any of the modern day desktop microprocessors. With a 10 cycle L2 latency, any application that fits within the Pentium M's 2MB cache will most definitely perform very well on the CPU. It is the 10 cycle L2 that allows the Pentium M to be competitive with much higher clocked CPUs in most mobile applications as they are normally office application tasks that are generally very cache-friendly. Keep this in mind as we look at the actual performance numbers of the Pentium M.

Understanding Pentium M Architecture Memory Latency and Bandwidth
POST A COMMENT

76 Comments

View All Comments

  • Lupine - Wednesday, February 16, 2005 - link

    I'm surprised at these results. I'm setting up a new Dell Inspiron 9200 (M 725 @ 1.6GHz/400MHz FSB) and it is schooling both my Barton 2500+ @ 2.2GHz and TBred B 1700+ @ 2.2GHz running Stanford's Folding@Home project (600 point proteins: ~37min per frame for the XP boxes compared to ~34min per frame w/ the laptop).

    So, if it is so weak, what is allowing it to process WUs at such a competitive rate? Sure, that is slower than an A64, but competitive w/ most P4 procs.
    Reply
  • fitten - Thursday, February 10, 2005 - link

    Something else to remember about the Banias/Dothan line of chips... Agressive power reduction was the #1 goal of the design process. In a 'normal' chip design, not all pipeline stages are the same length, the clock speed it runs at is the speed of the slowest part of the CPU. Since power usage is directly related to the frequency of the switching gates, the Intel engineers actually deliberately slowed down some parts of the chip to match the target release speeds (or get close to them) to reduce power consumption. This is, perhaps, the main reason why the frequencies don't scale so well as some would want them to scale. Reply
  • Visual - Thursday, February 10, 2005 - link

    here's another thought... when the opterons launched initially at ECC DDR266, there were similar comments like "give it unbuffered DDR400 or higher and stay out of its way" :) well, now that we have that, ok it did improve performance a bit. but not hugely. shouldn't help the dothan significantly more too. Reply
  • Visual - Thursday, February 10, 2005 - link

    I like how AMD got beaten by the P-M :) not because im intel fan, just because this will make things more interesting now.

    don't catch flame from this comment :p its my oppinion

    Funny how you picked the game benchmarks btw, its almost as if you wanted to show the P-M lacking behind the A64... from what I've seen it beats A64 in HL2 and CSS, and that's a game you don't skip usually :) so why now?

    Also looks suspicious how in lots of tests where P-M performs well with the A64 clock-for-clock or beats it, there is almost no difference in the 3800+ and 4000+ results... like if L2 isnt all that important, yet L2 is exactly how everyone explains the P-M success

    Maybe we'll see some 2MB L2 A64 "emergency edition" once Dothan gets a decent desktop chipset, just like what intel did to (try to) save P4 from the A64 :)
    actually i'd be happy if Dothan motivates AMD to develop faster L2 cache or something.

    Knowing Intel, i dont expect they'd even try to match AMD's prices with the P-M... and there's a lot of room for AMD to decreace prices, as they're selling with quite a margin now. So for sure the P-M won't be cost-effective compared to A64, not if you don't care for ultra-low power consumption at least.

    also it doesn't look likely Dothan could scale beyond 2.6GHz on current 90nm tech. by the time it gets there, AMD should've launched the 2.8 FX and most likely 3GHz too. so I have no doubts AMD will keep the lead for quite a while... maybe the race to 65nm will be the next turning point, as it seems its going smooth for intel (at least for P-M)

    anyway, even if AMD is better in absolute performance, pricepoint and (arguably) clock-for-clock, you gotta admit it to the P-M, it does quite a punch. fun times are coming :)
    Reply
  • Zebo - Wednesday, February 09, 2005 - link

    dobwal buy intel if you want mhz, AMD is for performance. Reply
  • dobwal - Wednesday, February 09, 2005 - link

    i wasn't referring to the FX series. Plus you are not understanding the point i was trying to make. Lets take a look at the FX series.

    OPN Model Operating Freq. Package ADAFX55DEI5AS FX55 2600MHz 939-Pin
    ADAFX53DEP5AS FX53 2400MHz 939-Pin
    ADAFX53CEP5AT FX53 2400MHz 940-Pin
    ADAFX51CEP5AT FX51 2200MHz 940-Pin
    ADAFX51CEP5AK FX51 2200MHz 940-Pin

    the first FX51 was release around late third quarter 2003. So in a little over a year the FX series has only increased 400 Mhz. Can you automatically assume that the FX has poor scalability in terms of cpu speed. NO. You know why, because the EE is underperforming and can't touch the FX. AMD has no need to push large scale speed increases out of the FX line, which would do nothing but increase cost with each new stepping it used to boost performance.

    The same goes for the Dothan at 2.26Ghz by the end of 2005. What other cpu offers the same level of performance vs. battery life. So why push for performance except to push sales.

    You simply can't determine the scalabiltiy of a cpu based on its roadmap especially when its the performance leader in its market segment and has no current viable competitor or one in the near future.
    Reply
  • Aileur - Wednesday, February 09, 2005 - link

    Oh and, superpi relies on the fpu to do its calculations, so so much for this fpu is crap trend we have going here.

    http://mod.vr-zone.com.sg/Aopen_i855_review/25sPIm...
    Reply
  • Aileur - Wednesday, February 09, 2005 - link

    Oh and before you start bragging about the better superpi1mb result of the a64
    http://www.akiba-pc.com/DFI_855/d17g_2608_spi1m.gi...

    this is 1 sec better, with 100mhz less, and single channel ram.
    Reply
  • Aileur - Wednesday, February 09, 2005 - link

    Since you seem to like xtremesystems
    http://www.akiba-pc.com/DFI_855/d15g_2435_spi1m.PN...
    also a 1ghz overclock, also on default voltage

    Id like to see how an a64 would perform on a kt266 (if that were possible)

    Give the pentium m time to mature and all those "OMG HAHA YOU CPUZ IS SO HOT LOLOL!!!1111" will be obsoleet.
    Reply
  • Zebo - Wednesday, February 09, 2005 - link

    58 "How long has A64 been stuck on 2.4Ghz."
    ----------------------------------

    There not. 2.6 FX-55 been out for months. More importantly AMD does'nt have to release new chips the way they dominate the benchmarks now. Could they? Hell ya.They got a nice buffer going, New FX's hit 3.0 on stock air. Cheap 90nm's are now hitting 2.7 on default Vcore and air. And by air I mean AMD's cheap all aluminum HS with a itty bitty 15mmx70mm fan, not Prescotts copper core screamers.

    T8000- You're clueless. Maybe it's the heat generated by your prescott making your head woozy, I dunno, but have a look here..1800 Mhz to 2800 Mhz on default Vcore stock fan.
    http://www.xtremesystems.org/forums/showthread.php...
    Reply

Log in

Don't have an account? Sign up now