Pentium III 500 Platform Comparison

Pages: 1 2 3 4 5 6 7 8

Memory Subsystem Performance

Memory performance has been the major focus of most publications and marketing departments recently, primarily because of the memory war going on between Rambus and competing architectures, as well as the rapidly increasing processor speeds. Since the memory subsystem includes both the system memory and cache, it can be difficult to separate them and see the true impact of different memory architectures. Fortunately, since both L1 and L2 cache reside on the processor, and all of these platforms use the same processor, this isn’t a problem here.

Unfortunately, Winbench99 does not include a benchmark test specifically for memory performance, however Membench is well-suited to the task because no special drivers are necessary to achieve optimal memory throughput. Note, however, that when multiple tasks are running concurrently, there could be a big difference in performance from the results shown here. As a general rule, single applications do not utilize the full bandwidth of these platforms, however multiple concurrent applications could easily saturate the memory bus. Both Passmark Performance Test and Membench are single-tasking benchmarks, so the potential benefits of DRDRAM cannot be measured here.

One additional benchmark, called Cachemem, was also used, which runs under DOS as well. This test measures the bandwidth of cache and system memory, as well as the latency and bandwidth of system memory using various data block sizes. The results are not included here, because I was unable to identify the author and could not verify the accuracy of the test before publication. For those who might be interested in seeing the raw numbers, I have included them as an ‘addendum’ to this report, without any accompanying analysis.

Passmark

AX6C/DRDRAM

AX6BC/SDRAM

AX63/SDRAM

AX63/VCSDRAM

Alloc Small Block

115.5

117.3

125.9

117.1

Cached Read

66.4

67.6

67.4

67.2

Uncached Read

50.3

52.9

54.7

54.6

Write

39.3

40.1

40.2

40.4

Memory Mark

65.4

66.9

69.4

67.2

Membench

AX6C/DRDRAM

AX6BC/SDRAM

AX63/SDRAM

AX63/VCSDRAM

Read Rate

266.66

333.33

327.86

333.33

Random Read Rate

111.73

125.0

104.71

121.21

Write Rate

246.91

183.48

246.91

243.90

Random Write Rate

48.54

53.61

39.13

33.38

Transfer Rate

141.84

126.58

150.37

166.66

Although the Apollo Pro133 chipset has been reviled in a number of publications as having ‘poor’ memory performance, it is apparent here that it isn’t as bad as some reports have indicated. It is also interesting to see that the i820 chipset has very poor performance with the exception of the Membench ‘write rate’ test results.

This illustrates the very trap that many well-meaning ‘review’ sites fall into by depending solely upon application level benchmarks. In almost every case, performance differences between the various chipsets have been attributed entirely to the memory performance, with the Pro133 chipset being singled out as having bad performance in this area. As can be seen here, the difference is not that great in most cases, and in a few is actually much better. It also appears that while the Pro133 supports VCSDRAM, it isn’t implemented very well, as it should in no way have worse performance than SDRAM on the same board. It will be interesting to see if this has been corrected in the Pro133A or KX133 chipsets.


Pages: « Prev   1 2 3 4 5 6 7 8   Next »

Be the first to discuss this article!