A member here, pwnzor, tested this many, many years ago with the help of a number of other great members, such as spfoam1, hellcamino, and russianhaxor. Most of the testing was with DDR2 memory, but there was some DDR3 memory testing. Part of the conversation was also about linking and synching the memory ratio to the FSB, as it was back in the glory days of Socket 775. A lot of the testing was in performance in 3DMark06, and
did some testing using games to see framerate differences for real-world performance. I think a fair set of conclusions are as follows:
- Although the testing found differences, the amount of difference wasn't very high. Less than a couple hundred points in 3DMark06, and relatively 1-2FPS in gameplay. (Testing was done between 700-1200mhz or so, with everything from 700mhz CAS 3 at 1T, to 1200mhz at Cas 5 at 2T; a couple people threw in DDR3 at 1333mhz)
- 1T to 2T command rate timings make a far bigger impact than the rest of the latencies (at least for DDR2. I make no claims for DDR3).
- Faster memory speeds will generally beat out lower latencies; in the case of low latencies using 1T against faster memory with higher latencies at 2T, there is a performance penalty until there is a significant difference in memory speeds to make up for it (an average quoted was ~300mhz for DDR2).
- As CPU's are overclocked, as the FSB is increased, and as the PCI-e bus is overclocked, the gap opens wider between higher-speed memory and lower-latency memory. One member noted that his 680i FSB clearly bottlenecked memory, as his X38 would see performance gains after the memory was pushed increasingly faster than the FSB (not to mention the way Intel chipsets would overclock the FSB much higher).
With regard to the fourth point, that may suggest that the memory was already being bottlenecked by other components in the system. While performance will go up based on the speed of all components, merely raising the speed of the memory had less effect. This is why there was such a performance penalty for running past linked/synched with the FSB or running 2T timings, compared to 1T. If the memory was bottlenecked, then it would make more sense that tightening the timings would make a larger difference in performance than raising the clocks speeds.
One thing that members couldn't have anticipated at the time of the thread was how well Intel's choice of going back to an on-die memory controller would improve memory performance, nor the advances in graphics cards and CPU's that can take advantage of more bandwidth. At the time, DDR3 was still relatively new, and being run with the technology at the time. To that effect, we're talking many people still had 680i's, some had 780i's, others were using X38's. At the time, it was Conroe/Kentsfield, with the only Yorkfield around was the QX9650. Although several members hinted at the possibility that bandwidth limit was a concern, it's more apparent now how much hardware has changed.
As mentioned earlier, the focus of the thread was in consideration more of system benchmarking and gaming, rather than a bunch of SiSoft Sandra or Everest tests for sheer memory performance. It's important to note that one of Intel's main advancement on the high-end desktop segment has been to go from Dual-Channel, to Triple-Channel, to Quad-Channel memory. While there are important performance gains from other components and engineering, it's interesting how Intel's continuing to push for higher performance through memory bandwidth. Additionally, the layout of the X79 boards was designed to also limit latency to the processor from the various channels. I think with the newer hardware, and the direction it's going, I can make a couple more observations based on the testing data:
- The DDR2 testing results in the aforementioned tests are limited in value (to today's hardware) because of the smaller amount of memory bandwidth needed at the time, as well as the bottleneck created by the FSB and ratios needed to run at certain speeds.
- The importance of lower latencies diminishes as the system can utilize more memory bandwidth; raw speed becomes far more important.
- The hardware available today can utilize memory far more effectively and efficiently than the hardware available when those tests were initially run, which would have a significant impact on the results of the earlier test.
Of course, lower latencies are better in comparison to higher latencies. Between a 1600mhz kit and an 1866mhz kit, I would not expect to see much of a difference if the latencies are similar. At 1600mhz to 2133mhz, however, it becomes a much different matter.
Essentially, the initial conclusion for the tests was that there's definitely a preference of either super-low and tight timings or higher-speed memory with good timings. Either way, you could purchase those sorts of kits for very cheap, without spending $500 for a set of DDR2 Corsair 1250mhz at CAS 5 timings (yup, we old timers). Moreover, FPS gains were relatively within 1-2FPS regardless of speed. Later benchmarking showed that higher memory speed trumped lower timings once the speed was high enough. The differences were still relatively small, but there was no question which was better.
What you should consider now is going for a good all-around kit that has both high-speed and good (but not necessarily very-low) timings. Err on the side of higher speed, rather than lower latency, as your CPU and the rest of the hardware rely on more bandwidth than they do clock cycles. There are a lot of of good kits out there, and you don't need to spend a lot to get what will work best for your system. Ultimately, that was the point of the thread in the first place; you can run higher clocked memory with slightly higher latencies, or slower kits (or slow down a faster kit intentionally) with very low latencies and not pay out the nose for either kit. This allows you to save money to spend on other hardware.
Of course, I'd be happy to see someone do a lot more testing on their own, and would be just as happy to throw a Blue Ribbon their way, as well. *cough*