Besides the practical study of the memory modules overclocking potential with different settings, I couldn’t help paying special attention to checking out their performance. It is extremely interesting to see how big of a performance boost will the system get if the memory is running at 1150MHz or 1250MHz. It is evident that the bandwidth of this memory subsystem is much higher than the processor bus bandwidth. Therefore, the question is: how justified is the intention of the overclocker memory makers to continue pushing up the working frequencies of their solutions? Especially, since lowering the latencies may eventually be a more efficient way of boosting the performance than raising the frequencies.
In order to check the efficiency of high-frequency memory I compared the performance of systems using DDR2-1150 and DDR-1250 SDRAM against the performance of systems equipped with DDR2-800 and DDR2-1067 with common timing settings. Note that among the results are the numbers obtained from memory modules with extremely low latencies. I am talking about DDR2-800 SDRAM running with 3-4-3-6-1T timings. Although the overclocker memory module makers do not officially ship solutions like that to the market, high-quality memory using Micron chips can sometimes work with these aggressive latencies. For example, for our today’s tests in this work mode I took Corsair Dominator TWIN2X2048-8888C4DF kit.
The first round of tests was performed in a system with 266MHz nominal FSB frequency. The testbed was absolutely identical to the one I have already described above. The only difference was the Core 2 Extreme X6800 CPU that was running at 3.2GHz frequency obtained as 12 x 266MHz.
First of all take a look at the practical bandwidth and memory subsystem latency measurements obtained with Everest utility:
The numbers are quite illustrative. The measures bandwidth of the bus between the processor and the memory reaches its maximum with DDR2-800 SDRAM with aggressive timings. Strange as it might seem, but the high-speed DDR2-1150 and DDR2-1250 loses here. So, it turns out that the bandwidth of the dual-channel DDR2-800 SDRAM is quite sufficient for efficient data exchange with the processor along the system bus working in nominal mode. In fact, this is quite logical. The bandwidth of the Quad Pumped processor bus working at 1067MHz is 8.5GB/s, while dual-channel DDR2-800 SDRAM should theoretically be able to transfer data at 12.8GB/s.
However, if we measure the latency of the memory subsystem, both DDR2-1150 and DDR2-1250 SDRAM, show better results than DDR2-800 memory with aggressive timings. From the practical latency standpoint, DDR2-800 with 3-4-3-6-1T timings loses to DDR2-1067 with 4-4-3-11-2T timings. It makes us question the undisputed efficiency of the DDR2-800 SDRAM with low timings, and hence we simply cannot make any conclusions without running some tests in real applications.
As we see, the situation in benchmarks aimed at revealing the performance of the entire system as a whole is somewhat different from what we have just seen in synthetic Everest tests. High-speed memory working at 1150MHz and 1250MHz allows gaining some noticeable performance advantage.