DDR2 needs to evolve as high-end memory

DDR2 IS INCHING close to the 600MHz clock /1200MHz DDR throughput/PC9600 grading - let's use the correct description after a complaint from a Micron guy - but are all these complex looking heat sink and cooling systems really helping you crank up the last MHz?

The argument is that, to overclock more on the bandwidth side, you have to up the voltage, and so increase the heat output of the DIMMs - and yes, huge, finned heat spreaders or liquid cooling or whatever are needed to take away that heat.

How much do they really help? Early this year, I had a 2x1GB set of Micron DIMMs, declared DDR2-667CL4 - standard fare without any heat spreader. There was also Crucial Ballistix DIMM set, same 2x1GB, with full heat spreader and, supposedly, same declared speed and chip type (both were out of the fabs in the same 4Q 2005). Guess what? both overclocked to DDR2-800 CL 4-3-3-10 at 1.9 volts easily, but higher voltages of 2.0 and 2.1 volts didn't help get better latency. Of course, the Intel 975X couldn't exactly do more than DDR2-800 on an Intel board anyway, so I couldn't push the bandwidth further.

In that case, the presence of heat spreader didn't help at all to get better figures. The heat? Well, after running the whole day, the chips were reasonably warm, but nothing like trying to touch a northbridge without a heatsink...

We are also looking at how the new Corsair and OCZ DIMMs, with enourmous heat sinks, seem to hit the performance wall despite all the tweaks, bells and whistles. And these are two top-notch vendors in this field. Then, this past week in Taipei, we looked at new DIMMs from local overclocker memory heavyweights, GEIL and G.Skill (don't confuse them, these are separate companies!), and found some really interesting declared latency settings, with high voltage to boot.

How does G.Skill CL 4-4-4-5 at DDR2-1066 sound at 2.4 volts? Or, furthermore, CL 4-4-4-12 at DDR2-1160 on the GEIL DIMM (but at 2.45 volts)? These are declared latency settings, printed on the DIMM itself, and therefore fall under the vendor's long-term warranty. As you can see, both of these are using pretty much the standard heat spreader - it has holes on top so that, if I used as a memory DIMM cooler fan set, it could blow air directly UNDER the heat spreader towards the dies, not just above the spreader surface. We'll see how they both perform in reality, this Xmas weekend.

As mentioned before, DDR2 is being driven beyond its limits, with these high voltages potentially affecting the lifetime of the modules - unless... the vendor has an an advantage in getting a good die selection from the fab (i.e. good latency & speed selection, and good voltage tolerance), and PCB design can take care of extra power and heat drive.

As Taiwanese vendors go into the same high-end enthusiast memory markets long coveted by the likes of Corsair, OCZ and Crucial, they will not be able to establish strong branding just by following the incumbent's approach. New approaches in getting hold of better dies with finer speed sorting process, as well as further PCB design fine tuning, could be just as worthy as, say, designing a HP-inspired micro water-cooling device where a miniature inkjet print hear spits 100's microdroplets of coolant instead of ink - on every DRAM die (the benefit is immediate evaporation and better die coverage / dissipation efficiency, though).

Our feeling is that, in reaching the 'DDR2 overclocker's holy grail' of PC9600 (DDR2-1200) stable operation, these two factors will play a far more important role than a complicated cooling system - in fact, if you got the right dies, you might just as well leave that high-end memory module open and simply provide good airflow with a couple of fans over it... µ