System memory bandwidth, which is often measured by the amount of data that can be transferred between system memory and the central processor, is similar in concept to graphics memory bandwidth. With processors increasing in speed at unprecedented rates, reductions in relative memory bandwidth throughput and increased latency times are forming a bottleneck in the performance of high-end applications such as 3D gaming. Fortunately there are a variety of memory bandwidth saving techniques that developers can employ which have been well documented by AMD and NVIDIA.

One of Sandra's tests is measuring sustained memory bandwidth by dynamically generating data and populating between 40-60% of available physical system RAM. Similar to the processor tests, Sandra utilizes processor specific optimizations as well as data pre-fetching techniques. All memory bandwidth tests were conducted with two memory timings (Row Active Delay - RAS To CAS Delay - Row Pre-Charge Delay), which were configured in the BIOS - the default of 6-3-3 and the optimized default 4-2-2 setting. The default BIOS setting for CAS latency is 2.0. Note that the 6-3-3 memory timing was only used for memory bandwidth testing.

Sandra's memory bandwidth test on the Athlon XP 1800+ / nForce / 256MB PC2700 system were 1862 MB/s for the integer test and 1805 MB/s for the floating point test, which is an increase of 38.7% and 37.1% respectively (based on the 4-2-2 memory timing).

The overall result of 4496 in the PCMark2002 memory test places this system near the top when compared against Athlon XP based systems with similar CPU test results.

The 4-2-2 memory timing results in a 1% increase in performance under the fastest Quake 3 graphics setting. Since games on high-end systems are being played at resolutions greater than 1024x768, tweaking the system memory will only provide a minimal gain in performance. At this point, the graphics sub-system begins to limit performance.

With the nForce supporting the ATA-133 interface, I outfitted the review system with an 80GB Maxtor DiamondMax Plus. The IDE interface continues to undergo incremental increases in data throughput and based on my testing, ATA-133 generally shows around a 20% gain over ATA-100. The following results using Sandra's hard drive test are based on a comparison of the primary ATA-133 drive against the secondary Maxtor ATA-100 hard drive with the nForce IDE Miniport driver installed. Maxtor's Accoustic Managment Utility was used to set each hard drive to performance mode.

The next two charts show the performance of the nForce IDE Miniport driver compared to the standard IDE driver that ships with Windows XP.

The one issue I had with the IDE Miniport driver is that once installed, it took Windows XP significantly longer to load. I haven't reported this to NVIDIA, but with the Windows XP IDE driver, the blue bar indicator on Windows splash screen would go through 5 cycles before transferring control to the logon screen. With the IDE Miniport driver, the bar indicator went through 26 cycles.

As I mentioned in the BIOS settings notes, there was a time during my testing when Windows XP failed to load using the IDE Miniport driver because I installed the operating system with APIC, or Advanced Programmable Interrupt Controller, mode enabled. The APIC is used to extend the number of available IRQ's available in order to avoid sharing and conflicts.

The review system arrived with APIC mode disabled, but the default BIOS setting is APIC mode enabled. To make a long story short, there was a point early on where I manually discharged the BIOS using a jumper on the motherboard at which time, and unknown to me, APIC mode became enabled. A subsequent install of Windows XP lead to the IDE Miniport driver issue, which I reported to NVIDIA. A few hours later I learned what went wrong and, unfortunately for me, it became necessary to re-run all the benchmarks I had planned on using. My original intention was to include much more material in this article, but I lost two valuable days because of an obsession with overclocking :)