Test Setup and Procedures
The art of performance tuning involves the identification of where the biggest bottleneck is and eliminating it, then proceeding to the next bottleneck. With overclocking there are only three potential bottlenecks which can be addressed, namely the speed of the processor, the memory subsystem and the I/O subsystem, with the overall effect on performance depending upon the specific circumstances. The intent behind these tests is to isolate each of the three elements that comprise a computer system and evaluate the performance impact, if any. Other factors, such as the speed of the actual devices themselves (cache, hard drive, memory, etc.) are not affected by, and may in fact limit the effects of overclocking.
Unfortunately, the majority of people seem to see overclocking as an end in itself, rather than as a means to achieve greater throughput. Though all who engage in the activity will claim that their intent is to have the ‘fastest’ system, in many cases there seems to be little understanding about what bottleneck is being addressed. Some simply try to push their processor to it’s maximum speed (MHz) without considering whether it is already starving for data. Others may increase the memory or PCI bus speed because they believe that faster must mean better performance. Note that increasing the memory or I/O bus speeds is in effect increasing the bandwidth. If that element is not ‘bandwidth constrained’ there will be little or no benefit, and yet may reduce the stability and reliability of the system.
The specific benchmark numbers presented should not be compared to those tossed around on the newsgroups or on other web sites, as they will very likely not be the highest you have ever seen. The intent is to show the relative difference from a ‘baseline’ that various changes to the processor, memory and I/O speeds have on different benchmark tests. Of particular note is the frame rates for Quake, which will be very low compared to what most gamers report for their tests. This is primarily because of the specific video card used, as well as the lack of any special drivers or components for increasing video output speed. To reiterate, the intention here is to show how changing the speed of each element may or may not have an effect on throughput depending upon the situation.
Four common test suites were chosen to run under three operating systems. For Windows 95, the Winstone 98 Business test suite was used. For Windows NT, both Business and High End Winstone 98 test suites were used and under DOS the Quake 1.06 Demo2 test was used. There are numerous other benchmarks which could have been used, however these test suites provide the best selection of commonly used applications, and are all available to anyone for free. This makes the results repeatable by anyone who wishes to generate their own results as further proof (or contradiction) of the conclusions drawn here
An attempt was made to keep as many of the parameters the same as possible. The same video card and hard drive were used for every test. For both Socket 7 and Pentium II tests, the same PC100 SDRAM was used,while the Pentium Pro tests were run with 60ns EDO. The major difference to be aware of is the L2 cache size and speed, which ranged from 256K at full speed (Pentium Pro) to 512K at half speed (Pentium II) to 512K and 1MB at system bus speed (Socket 7). The tests which isolate the PCI bus speed were extremely simple at 75MHz, but were much more of a problem at 83MHz. In the former case, the same motherboard was able to switch from synchronous operation to async, however there are literally no motherboards with this capability at 83MHz. In order to test the effect of the 41.5MHz PCI bus (one half of 83MHz), two boards were used and compared both at 66MHz and at 83MHz bus speeds.
The processors and motherboards chosen for these tests were the AMD K6-2 / AOpen AX59Pro, Intel Pentium II (Klamath) / Abit BX6, and the Intel Pentim Pro / M Tech R651. For the PCI tests, the M Tech R581A was used for both sync and async operation at 75MHz, while the R581A and AOpen AX5T were used for the 83MHz tests. The main issue here is the difference in L2 cache size, which was 1MB for the R581A, but only 512K for the AX5T. The benchmark scores at 66MHz were used to establish a basis for comparison at 83MHz. Again, all Socket 7 tests used the same K6-2 processor and PC100 SDRAM.
Be the first to discuss this article!