One of the many issues presented with a SO-DIMM capable system, whether laptop or desktop, is one of performance.  In our recent Haswell memory scaling article using regular sized DIMMs, the high-performance sweet spot for memory was around the 2133 MHz CAS 9 or 2400 MHz CAS 10 marks.  The issue with SO-DIMM systems is that memory often starts at 1333 CAS 9 or 1600 CAS 11, but in recent months companies like G.Skill, Corsair and Kingston have released higher specification SO-DIMM kits, up to 2133 CAS 11.  This is still a little way off our sweet spot, but on the road.  The main barrier to this incidentally is the lack of XMP support on laptops and mobile devices, firmly shutting the door on speeds above 1600 MHz without a modified BIOS.

While we are doing some in-house memory scaling testing regarding SO-DIMM, G.Skill went ahead with some testing using the main overclockable motherboard for SO-DIMMs: the ASRock M8.  We reviewed the ASRock M8 as a Steam Box alternative last year capable of handling an i7-4770 CPU and a 250W GPU and gave it a Silver Award for industrial design. 

For the overclocking test, G.Skill use their DDR3L-2133 MHz CAS 11 2x4GB 1.35V memory kit and boost the final speed to DDR3-2600 12-14-14.  Back in our memory scaling article we introduced the concept of a memory Performance Index as a rough guide to performance, and this memory kit started at a PI of 193 and ended on 217, or a 12.4% increase in potential performance.

While G.Skill have jumped us in terms of showing that these speeds are possible, it remains to see if memory manufacturers will go ahead and make SO-DIMM modules at this speed. Or ultimately what matters more is that the platforms that use them (especially laptops and SFF) will actually adhere to XMP and allow us to enable it without any fuss.  There are speed gains to be had by moving up from the industry default of 1600 MHz CAS 11 as we showed in our Haswell memory scaling article, but there needs to be a paradigm shift from the manufacturers that implement SO-DIMM.  If the SO-DIMM modules come up to par with regular DIMMs, there might be a future where motherboard memory makes the transition. 


Comments Locked


View All Comments

  • TiGr1982 - Friday, February 21, 2014 - link

    Laptop Haswell uses 1.35 V DDR3L, if I'm not mistaken.
    Most of SO-DIMMs are for laptops, and most laptops are now Haswell.
    So, IMHO, this initiative with going much beyond JEDEC DDR3 1600 for DDR3 SO-DIMM's, which requires raised voltage for DRAM, won't really get a wide market adoption.
  • Ian Cutress - Friday, February 21, 2014 - link

    The modules used for this were 1.35V 2133 MHz. That means 2133 MHz at 1.35V. If you bin the ICs correctly, then eventually you it enough which have the speed/voltage characteristics you need.
  • TiGr1982 - Friday, February 21, 2014 - link

    OK, indeed, this DDR3 2133 SO-DIMM works at 1.35V, being actually DDR3L 2133, and does not cost like a jet plane - I overlooked that initially (because I haven't seen any desktop DDD3L 2133 1.35 V offerings). Thanks.
    Now, SO-DIMM systems vendors should support this speed in BIOS - then it can get some traction.
  • TiGr1982 - Friday, February 21, 2014 - link

    Enthusiast-oriented DRAM module manufacturers always tend to push premium higher frequency and/or lower timings modules, because these are more expensive than the regular ones. But these usually require higher voltage, as I said, so they usually remain niche offerings. The rest of the industry continues to use the standard JEDEC modules.
  • Ian Cutress - Friday, February 21, 2014 - link

    As pointed out in our Memory Scaling review, there is a law of diminishing returns with faster memory. But JEDEC standard 1600 CAS 11 is a hole of performance loss - up to 20% in some benchmarks over 2400 CAS 9, which by any stretch is becoming mid range vs. the 3000 MHz memory kits available.
  • TiGr1982 - Friday, February 21, 2014 - link

    That's right - I'm running two G.SKILL RipjawsX 2400 F3-2400C11D-16GXM kits on my Haswell Core i7 desktop myself (32 GB total), but at required 1.65 V though (bought these two kits last summer for $150 each, not for the current $200 price). For me, $150 made sense for 2X8 GB DDR 2400 kit (even CL11).

    Technically, indeed, you can buy DDR3 up to 3100. However, price-wise, I would not really consider anything beyond 2400, because DDR3 prices skyrocket ridiculously fast after 2400 - this should be the result of the aforementioned binning.
    So, I would say, the reasonable ceiling for desktop DDR3 in terms of street price currently is 2400-2600, and not 3000-3100. For SO-DIMMs this ceiling is now 2133 (I'm counting without overclock - at MFG rated speeds).
  • Khenglish - Friday, February 21, 2014 - link

    This G.skill is only single rank, so it has bandwidth as if it were 2133.

    The fastest memory for laptops is still 3 year old samsung M471B5273DH0, which can do 2400 11-13-13, or 2133 9-10-10, and it's dual rank.

    As for how the g.skill does at 1.5V vs 1.35V... it doesn't help at all for some reason on my p150em with the 8GB dual rank version (I cannot get 2133 to run at any timings in dual channel).
  • MikeMurphy - Friday, February 21, 2014 - link

    In this day and age the motherboard can't pull appropriate memory timings from a table based on a set voltage?

    I hope JEDEC has come up with a better system for DDR4 timings, because this is inexcusable.
  • Khenglish - Friday, February 21, 2014 - link

    It'd be fine if more laptops supported XMP. As far as I know only the m18x R2 fully supports XMP, with only a handful of other laptops even having partial support.
  • TiGr1982 - Saturday, February 22, 2014 - link

    Indeed, while lots of "above average" desktop boards do support XMP, usual "above average" $1000-1500 laptops mostly do not. That's an issue preventing the distributio of fast SO-DIMMs.

Log in

Don't have an account? Sign up now