AM2 Memory Performance

The move by AMD from the current Socket 939 to Socket AM2 is pretty straightforward. We know the new AM2 processors will continue to be built using the same 90nm manufacturing process currently used for Athlon 64 processors; AMD does not show roadmaps with AM2 processors built on 65nm until early 2007. To this point AMD has also reiterated that AM2 will not bring any changes to the Athlon 64 core. In other words, the socket will change to the new AM2 Socket 940, but under the hood the current 939 and the upcoming AM2 940 will beat with the same heart. The only substantive difference expected with AM2 is the move from DDR memory to official AMD DDR2 Memory support.

With that in mind it is time to delve more deeply into the what is really new in AM2 - support for DDR2 memory with AMD's unique on-processor memory controller. Many have expressed expectations of remarkable performance increases for DDR2 on AM2. This would be at odds with what we have seen from DDR2 in the past. With the move of Intel's NetBurst architecture to DDR2 there were really no gains at all in memory performance. Those expecting big gains point out that the AM2 on-chip memory controller, like the Athlon 64 on-chip DDR controller, should provide much lower latency and higher efficiency than Intel's chipset-based memory controller for DDR2.

This is our first opportunity to look more closely at an AM2 DDR2 controller that might answer these questions about memory performance, since it is the first AM2 design to outperform Socket 939. Earlier AM2 spins could not match 939 memory performance, but they continued to improve. This is remarkable when you consider that new Intel processors pretty much have performed like final shipping processors some 5 months ahead of launch. AMD, on the other hand, has done most of their development work on the DDR2 memory controller in the last 3 months with just 6 weeks remaining before launch.

The most recent AM2 roadmap is still showing AM2 launching June 6, 2006 at Computex in Taipei. With just 6 weeks to go before launch, there is not a lot of time for surprises with AM2. As pointed out in AMD Socket-AM2 Performance Preview, there is not much wiggle room when OEMs expect mid-May shipments of AM2. All of this leads us to believe that our fourth spin of AM2 this year is very close to what will actually be shipping on June 6th. We can always hope for surprises, but given what AMD has said so far we should be very close to final silicon.

You already know that the AM2 does modestly outperform Athlon64 Socket 939. What will be explored here is how the memory controllers compare in latency and bandwidth, memory performance at various DDR2 settings compared to fast DDR400 2-2-2 memory, and basic overclocking performance of AM2 compared to Socket 939 when the CPU and memory are both pushed to improve performance.

Memory Test Configuration


View All Comments

  • peternelson - Saturday, April 15, 2006 - link

    I notice all your tests were performed in 32 bit mode.

    This cpu can handle 64 bit instructions.

    While 64 bit registers (and more of them) allows faster data manipulations, that advantage is traditionally offset by the need for bigger wordsize of instructions.

    So if the memory reading of the instructions is more memory-hungry that could be more use for this extra memory bandwidth.

    Therefore I suspect IN 64 BIT MODE, there could be more advantage on a fast DDR2 than on a bandwidth-limited DDR system.

    How to test this? Well you could run some 64 bit Windows and BENCHMARK FAR CRY in 64 bits version as it is available as 32 and 64 bit.

    See if running in 64 bit with this new ddr2 memory negates the disadvantage of limited bandwidth for instruction feeding?

    If so this would be increasingly an advantage in future as more people move to 64 bit OS, including at Vista-time.
  • smitty3268 - Saturday, April 15, 2006 - link

    The larger instruction size is barely an issue. The real difference is that all pointers are doubles in size from 32 to 64 bits. This can lead to a significantly lower number of variables stored on the cache, which can lead to increased bandwidth usage. Reply
  • AnandThenMan - Saturday, April 15, 2006 - link

    I agree with what you are saying. In 64 bit mode, that A64 *should* benefit from the increased DDR bandwidth.

    The problem is, the 64 bit version of Farcry was basically a scam and offered no performance or visual increases solely because it was a 64 bit optimized game. If I remember correctly, the extra visual effects in the 64 bit version were basicially enabled if run in 64 bit mode, but had little or nothing to do with actually being optimized for the A64 in 64 bit mode.
  • peternelson - Saturday, April 15, 2006 - link

    I see your point too.

    Yes some additional features could have been done on 32 bit version but were restricted to 64 bit platforms. But surely the compiled binary was actually a 64 bit binary even if not optimised much? In which case it would still be interesting to compare.

    32 bit DDR versus 64 bit DDR and versus 32 bit DDR2 and versus 64 bit DDR2.

    My hypothesis is that the speedup (even if small) from using the 64 bit binary over the 32 will be greater on AM2 DDR2 than the same test on 939 DDR.

    I agree that Far Cry was not the best example, but you may know of other good benchmarks or games which are tuned for this.

    eg the same four tests of PRIME95 ( which is available in 32 and 64 bit. The Trial factoring test benchmark shows a good speedup in 64 over 32 bit operations. But then that doesn't use main memory much as it is highly optimised to work inside the L1/L2 cache. There must be other suitable tests though to compare 32 and 64 bit on some memory intensive task with binaries optimised for each architecture.
  • IntelUser2000 - Saturday, April 15, 2006 - link


    My hypothesis is that the speedup (even if small) from using the 64 bit binary over the 32 will be greater on AM2 DDR2 than the same test on 939 DDR.

    Right, the hypothesis for higher clock speed giving better increases were similar, however it gave less increases.

    Wesley, there is another typo. On this page:">

    DDR400 to DDR2-800 performance increase in CoD2 is said to be 10.6%. That is not correct. DDR400 to DDR2-533 is 10.6%, but DDR400 to DDR2-800 is only 6.7%. Check your calculation numbers please.
  • Wesley Fink - Saturday, April 15, 2006 - link

    The calculation has been corrected. Thank you for catching this and bringing it to our attention. Reply
  • peternelson - Saturday, April 15, 2006 - link

    You should not claim AMD's on-processor memory controller is UNIQUE".

    Unique means nobody else does it and it is a unique feature of AMD.

    THAT is incorrect.

    Although Intel don't do it, there are other chips that have on-chip DDR or DDR2 controllers including Clearspeed. I can even put a ddr or ddr2 controller (or several) into my own chip designs in a Xilinx FPGA because Xilinx license the design free for use in their chips.
  • Griswold - Saturday, April 15, 2006 - link

    Pretty unique in the x86 world, isnt it? Reply
  • peternelson - Saturday, April 15, 2006 - link

    No, actually it isn't. That was precisely my point.

    Transmeta Efficeon and VIA C7 can both have on die memory controllers too.

    They run x86 instructions quite happily.

    Don't get me wrong, it's a good idea, it's just not UNIQUE any more.
  • Wesley Fink - Saturday, April 15, 2006 - link

    Transmeta and the VIA C7 aren't really AM2 and Conroe competitors in most situations. However, I can conceive some applications where they might be. To be more precise I will try to use another word to describe the on-processor memory controller in the future.

    Do you work for VIA or Transmeta?

Log in

Don't have an account? Sign up now