Why 1GB DIMMs?

In past memory reviews, we have said a lot about the impact of the 2T Command Rate on memory performance. We have speculated with others that the Rev. E AMD processor would remove the 2T requirement for 4 DIMMs, but in fact, the release Rev. E still required a 2T Command Rate with 4 DIMMs on the new on-CPU memory controller. So, what is the impact of going from 1T with 2 DIMMs to 2T with 4 DIMMs?

Here, you see tests of 2 vs. 4 DIMMs of our standard OCZ PC3200 Platinum Rev. 2 based on Samsung TCCD memory chips. In all tests, the timings were 2-2-2-7 and the only difference was the 1T Command Rate with 2 DIMMs and 2T with 4 DIMMs:

Quake3 fps Sandra UNBuffered Sandra Standard Buffered Super PI 2M places
(time in sec)
Wolfenstein - Radar - Enemy Territory fps
2x512MB
2-2-2-7
2.6V 1T
550.2 INT 2621
FLT 2738
INT 5984
FLT 5938
80 121.1
4x512MB
2-2-2-7
2.6V 2T
529.6 INT 2276
FLT 2362
INT 4925
FLT 5938
83 116.9
1T Performance Increase 3.9% 15.5% 21.1% 3.8% 3.6%

For best comparison, we reran all benchmarks with the exact same settings for 2 DIMMs and 4 DIMMs, the only difference being 1T or 2T timings. We could have achieved faster timings (and a bigger performance difference) by tweaking each set of tests for best bandwidth. Therefore, these should be considered best case (lowest) performance differences.

As you can clearly see, the memory-only Sandra tests show a 21% drop in memory bandwidth in the standard buffered test when reducing to a 2T Command Rate. The more real-world Unbuffered Sandra test still shows a 15.5% bandwidth drop. Games and Super Pi, where memory is just one small component of the overall result, show about a 4% performance reduction. These represent the impact of this memory change on the overall system performance, where the Sandra scores measure impact on memory alone.

The second area where 4 DIMMs have a significant impact compared to 2 DIMMs is overclocking. This won't matter to those who run their systems at stock speeds, but it will be very important to those who try to squeeze the best performance from their memory. Here, we compared overclocking with our stock OCZ PC3200 Platinum Rev. 2 with 2 and 4 DIMMs. Again, we kept timings and settings exactly the same.

Highest Overclock at 9X Ratio
(4000+ Rev. E)
2x512MB
2.5-4-3-7 1T
318 (DDR636)
4x512MB
2.5-4-3-7 2T
240 (529.6)
2 DIMM Performance Increase 295%

Using the same DIMMs and same timings, we could only reach a 240 Processor Clock or +20% with 4 DIMMs. Two DIMMs allowed an overclock to 318 or +59%. With two DIMMs, we could overclock almost 3 times higher than with 4 DIMMs. You can likely overclock a bit higher with 4 DIMMs by further tweaking memory timings, but 4 DIMM overclocking will always fall far short of 2 DIMMs on current Athlon 64 memory controllers.

Higher performance at 1T vs. 2T and higher overclocking potential are certainly strong motivations to use 1GB DIMMs instead of 512Mb DIMMs when you want to populate an Athlon 64 board with 2GB. However, the gain of 1T in the past has often been more than offset by the slow memory timings available on 1GB DIMMs. The Athlon 64 controller is much lower latency than current Intel solutions, and thus, is much more bandwidth-sensitive to memory timings than a comparable Intel solution. We have shown in past memory reviews that slower memory timings can more than offset the 3.6% to 3.9% real-world performance loss that results in 4 DIMMs at 2T Command Rate. However, the availability of DDR400 1GB DIMMs rated as fast as 2-3-2 at DDR433 certainly has the potential to change this picture.

Index Impact of Athlon 64 Memory Controllers on 1GB DIMM Performance
POST A COMMENT

40 Comments

View All Comments

  • walmartshopper - Wednesday, October 12, 2005 - link

    Thanks for the review. I've been toying around with 4 sticks of Ballistix pc3200, only to get them running at 480mhz with 2.5-3-3-8 2T timings (I got 530mhz with 2 sticks at 1T). It's not too bad, but after reading this, I decided to replace them with the 2x1024 OCZ pc4000. I actually have 5 sticks of Ballistix, and I'm hoping to sell them for 50$ each. Anyone interested? Reply
  • AkumaX - Wednesday, October 12, 2005 - link

    I think on Page 2 when you were comparing 2T vs 1T you also meant 2x1024mb vs 4x512mb, rather than 2x512mb vs 4x512mb right? Reply
  • cryptonomicon - Wednesday, October 12, 2005 - link

    no, he's comparing the same ram to show how the memory controller goes to crap after you load up more than two dimms. the sticks were both 512x2 plat rev II Reply
  • TheInvincibleMustard - Wednesday, October 12, 2005 - link

    Nice, well-written article, Wesley. It seems slightly ironic to me, though, that this review comes out just as I'm expecting a 2GB kit from G.Skill via NewEgg to arrive tomorrow to replace my el-cheapo 2x512 (3-3-3-7 at DDR400? ick)

    Minor nit: pg 8
    "The performance differences will be that the NVIDIA 71.84 driver is a bit faster than the 61.77 and 71.84 drivers used in earlier memory reviews."
    Umm ... unless the driver is somehow faster than itself, I'm hoping that's a typo of sorts.

    -TIM
    Reply
  • n7 - Wednesday, October 12, 2005 - link

    I'm glad to see a review on 2x1024 MB; it was past due.

    I realize there aren't many DDR433 & up 2x1024MB RAM manufacturers, but i would have liked to have seen Crucial, Mushkin, Patriot, Geil, since they all make good DDR400 kits, & at least in Mushkin & Crucial's case, they also make DDR433 & up kits.

    To make it simple, i'd like to see a review with a few more companies involved :)
    Reply
  • rqle - Tuesday, October 11, 2005 - link

    Isnt it a little unfair to say one brand highest speed obtain @ 2-3-2-7 is lower then some other brand higest speed obtain is higher cause of 3-4-3-7? I mean, if you set corsair at a more lax timing AND higher voltage wouldnt it do better? Some one fill me in. Cause i remember back in the old days, memory i bought that can do 2-2-3-6 at 2.5v that was rated at 400DDR would overclock and do much better the same timing and voltage of some-old brand 533DDR+. Reply
  • Wesley Fink - Tuesday, October 11, 2005 - link

    From p. 5 of the review:

    "Increasing voltage beyond 2.7V did not allow us to go any higher in overclocking, nor did more relaxed timings allow us to push higher. The limit is DDR492 - very close to DDR500."

    We tried to go higher but DDR492 is the limit with the 1GB Corsair dimms we tested. As we stated in the review it is likely Corsair is using a different Infineon chip than OCZ and Gigaram, or they are binning for best performance in the DDR400 to DDR500 range. Gigaram and OCZ are probably also using different Inineon memory chips - or they are at the least using different binning methodologies.
    Reply
  • ozzimark - Tuesday, October 11, 2005 - link

    wow.. wesley, long needed article. however, some silly mistakes that i think i see :D
    first.. the gigaram oc'ing chart. max speed is put at 2-4-3-7.. are you sure it's not 3-4-3-7?
    also, the second speed is curiously 2-2.5-2.. where 2.5-3-2 is exepcted

    second, i know the difficulties of getting review samples, but where is the biggest name in 1gb sticks right now, crucial ballistix? i have seen many of these sticks do 280-300mhz.

    last, i remember the value ram overclocking article you guys had a while back. plan on going the same for 1gb sticks?
    Reply
  • Wesley Fink - Tuesday, October 11, 2005 - link

    I went back to my test logs and corrected the misplaced values. DD436 is 2-3-2-7 and DDR556 is 3-4-3-7. The second value in all reported strings is RAS to CAS Delay in case anyone is confused by the notation.

    We hope to do a review of the Crucial Ballistix 1GB dimms in the near future. We haven't decided whether to do a Value 1GB roundup yet, but we will consider your suggestion.
    Reply
  • Ender17 - Tuesday, October 11, 2005 - link

    ;) Reply

Log in

Don't have an account? Sign up now