Articles: Memory

Bookmark and Share

Pages: [ 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 ]


The Athlon 64 platform is highly sensitive to the memory subsystem latency. The memory controller integrated into the CPU doesn’t act as a bottleneck and this leads to a dramatic growth of the overall performance of the system as soon as the timings are lowered. A difference between 2-5-2-2 and 3-8-4-4 in the DDR400 mode amounts to 20% in average in real applications and is even bigger in some benchmarks.

So the notion that “modules with minimal timings are optimal for the Athlon 64” is still quite persuasive and is in fact true in most situations. But today we have an untypical case: few companies can offer DDR SDRAM capable of working at frequencies that high! We carried out a short test of the memory subsystem using the G.Skill F1-4800DSU2-1GBFR kit in default and extreme modes (the above-indicated frequencies at Vddr = 3.0V and 3.3V). To account for measurement errors we rounded the numbers off to 50KB/s.

The results were not quite what we had expected. Yes, the low latencies at DDR450 push the performance of the memory subsystem to the level of DDR650 with “slack” timings, but DDR400 with the extreme timings is considerably slower than DDR600 at the default timings. Yet this is all quite predictable. What’s much more exciting is the memory performance in the 2.5-8-3-3-1T @ DDR600 mode. It’s the first time we see a memory speed of nearly 9000KB/s in Sandra Memory Benchmark!

This outstanding performance is due to a single BIOS parameter: DRAM Command Rate set to 1T instead of 2T as is the default for DDR600. Alas, the beautiful number (DDR600) is not achievable in this case at 3.0V voltage (you have to increase it by 0.3V more), but we’d anyway like to compare the effect of DRAM Command Rate = 1 to an injection of nitro into a car engine. :)

Pages: [ 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 ]


Comments currently: 24
Discussion started: 08/20/05 09:58:44 PM
Latest comment: 08/25/06 12:43:21 PM

View comments

Add your Comment