Overview
Background and Platform DefinitionsIn this comparison, I attempt to identify the difference in performance between the ‘classic’ Athlon and the new Athlon-with-performance-enhancing-L2-cache (Tbird). The classic Athlon included 512KB L2 cache, and in this case, runs at 1/3 of the processor speed. The new Athlon has 256KB L2 cache, which runs at full speed. In addition, this cache is implemented as an ‘exclusive’ or ‘victim’ cache, which means that it does not mirror what is in L1, so the effective cache size is L1 + L2, which in this case makes the total cache size 320KB.
This platform comparison includes the following processors and motherboard combinations
- FIC AZ11 motherboard w/ KT133 chipset and TBird Athlon 900MHz (Socket A)
- AOpen AK72 motherboard w/KX133 chipset and classic Athlon 900MHz (Slot A)
Common components included:
- 128MB Crucial Technology PC133 SDRAM (3-2-2)
- 128MB NEC PC133 VCSDRAM (3-2-2)
- Diamond Viper V770 Ultra
- IBM Deskstar 22GXP 9.1GB UDMA66 IDE hard drive
- Toshiba XM-6602B 32x CDROM
- Sparkle 300W ATX power supply (model: ST-301HR)
The first thing I wanted to determine is what effect the different processor cache has upon the memory, graphics and I/O throughput. Most benchmarks do not test only the speed of the system memory, but also the cache itself. In this particular case, the chipsets used on each motherboard are almost identical, since the KT133 is basically a KX133 with some bug fixes and optimized for the Socket A configuration.
I decided to use ZDBop’s Winbench 99 and Passmark Software’s Performance Test for most of the subsystem performance testing. MadOnion’s 3Dmark2000 was used as an artificial, but standardized, game benchmark for those who might be interested. The DOS based Membench by Intelligent Firmware was also used to test memory and cache performance, as it is very reliable and consistent.
Following the component level benchmarks, I then ran the popular ZDBOp application level benchmarks to determine the ‘real world’ differences between the two cache implementations. These were performed under Windows 2000 rather than Windows 98, because of the additional information available from the High-End Winstone99 tests. The Content Creation and Business Winstone99 results are only provided as an overall average score, which has very little value in determining what types of applications benefit from different cache implementations.
As mentioned in the last platform comparison, benchmarks are less reliable when run under the Windows OS because of the additional layer between the software and hardware. The benefit of benchmarks which run under Windows is that the same drivers used for real applications are used by the benchmarks, providing a better ‘real world’ scenario. I had also indicated in that comparison (which Intel recently picked up on) that it is conceptually possible to have concurrent tests running to measure the actual bandwidth capabilities but that they do not exist at this time. The reason, of course, is that few, if any, users actually do any true multi-tasking on their desktop PC.
Using the various component level benchmarks, and comparing them to the application level benchmarks, we can attempt to determine what contributes to any differences that might exist, and therefore the reason certain applications might benefit from one platform or another. In all charts, the highest score is in blue, while the lowest score is in red. Every benchmark was run three times, rebooting between runs, with the given score an average of the three runs.
Be the first to discuss this article!