Core i7 high-speed memory runoff

WHEN THE NEHALEM integrated memory controller - seen in all its variations including Core i7 desktop and Xeon workstation and server parts - came out, hopes were high for a massive memory performance jump and near unlimited bandwidth benchmark scores.

Yes, we got the extra performance, but there was seemingly little - almost none in fact - incremental gain once we crossed the triple channel DDR3-1600 speed threshold. Even our colleagues like Anandtech confirmed that.

On top of that, the need to run the uncore CPU portion with the DRAM memory controller at twice the declared RAM throughput, that is, 4GHz for DDR3-2000, makes the overclocking work needed for this quite tedious as the uncore on Nehalem doesn't overclock anywhere nearly as well as the core.

Yet, for those willing to go all the way, there are top-bin triple-channel memory kits going to those stratospheric speeds, reaching DDR3-2133 and beyond. These kits, while costing more, also offer very good performance with low latency when running at lower speeds, and often at very low, power saving voltages too for extra reliability.

Here we had three new kits to compare: Kingston HyperX DDR3-2000 (yes, based on those Elpida "die-dies"), Geil DDR3-2133 and Gskill DDR3-1600 low latency. This time we focused on that "ideal performance" tri-channel DDR3-1600 setting with 3.2GHz uncore frequency for all these, and saw how far we could go in latency and voltage to run both Sandra bandwidth and latency tests repeatedly well. The test platform, as usual, was the Intel Core i7 975XE running at 4GHz on an Asus Rampage II Extreme mainboard, cooled by the CoolerMaster V10 TEC cooler and fed by a Corsair HX1000 high-end PSU.

The first to run was GEIL (Golden Emperor International Limited for those unfamiliar with the Taiwan brand) DDR3-2133 CL9 kit and, as you can see, after half an hour of tweaking, we got a very good - read Elpida class - set of latency marks: CL 6-6-6-15 at DDR3-1600 and at just 1.59 volts, passing all the Windows Sandra tests I ran over and over again, not bad! I also applied a GEIL Cyclone fan here, reducing the heat a bit but not much, as the DIMMs didn't run that hot in the first place. The funky LED self-promo round on the DIMMs may entertain you if the casing side door is transparent.

Next, the Kingston HyperX DDR3-2000 set came in. In fact, it was 2 sets, and luckily things were that way as one of the Elpida based DIMMs failed during the test. (I had used it for about a month before that). The other kit proved to be a very useful spare to complete the test, and up to now I've had no problems with it. It reached EXACTLY THE SAME latency settings as the GEIL, in fact even the same benchmark results in Sandra bandwidth and latency, except that it could run stable at an even lower voltage, 1.57 volts, at these settings, or at least the surviving modules could. Kingston also supplied the cooler, a less sexy but a bit more powerful dual fan model.

Finally, the Gskill DDR3-1600 CL6 kit, seen here compared to the fan-less Kingston Hyper X, ran fine at CL 6-7-6-16, about the same as the advertised 6-7-6-18 latency setting on the DIMM itself, and I needed 1.64 volts for it. So, a bit slower and a bit hotter than the rest. However, the large heat sink and, hopefully, smaller price should give this its niche.

Keep in mind that these are premium DIMMs and the price stratification is often done based on the real performance the DIMM can achieve. It's still way ahead of what good DDR3 DIMMs could do just half a year ago.

Here are the BIOS settings for the GEIL and Kingston DIMMs

and for the Gskill model - notice how the increased DIMM latencies affect the 2nd level memory channel roundtrip latency for each of three Core i7 channels in ns:

Then, take a look at the Sandra bandwidth performance...

as well as the latency benchmark on Geil & Kingston:

So, besides the most generous heat sink, the Kingston kit managed to match the top performance at a notch lower power consumption. Come to think of it, lower voltage might be of some benefit to lengthen the useful life of some of those Elpida dies.

The performance difference between the Geil and Kingston on one side, and Gskill on the other side, is minor. In Sandra 2009 bandwidth, it's roughly 1 per cent (28.5 GB/s vs 28.2 GB/s) and 4 per cent in latency (52 ns vs 54 ns for a 16 MB transfer start from memory). Compare that to the huge bandwidth result difference versus the normal DDR3-1600 CL8 memory or even DDR3-1866 CL9 memory, that completely loses out versus any of these despite its higher frequency. So, lower latency really offsets higher theoretical bandwidth. In fact, the price difference you obtain when shopping for these modules will be now as important as the latency finesse, as all of the modules are generally performing well.

While we look at the subsequent maximum frequency test as well (and it will depend on the particular mainboard a lot too), we should conclude here that, for now, DDR3-1600 CL6 low latency modules are the best stable performance choice for high speed Core i7 975XE users. When the first dual Core i7 (that is, Xeon W5590 and beyond) desktop and workstation profiled boards come out soon, I hope to see the bandwidth and latency support for this. That same DDR3-1600 CL6 across six channels of memory over two CPUs will be an interesting yet hopefully easily attainable performance proposition. µ