Athlon 900MHz Platform Comparison

Pages: 1 2 3

Subsystem Performance

Graphics Subsystem

I used two different benchmarks to specifically test the graphics performance: Winbench99 and Passmark Performance Test. Both of these run tests that attempt to isolate the graphics subsystem from the rest of the components, however there is still some reliance on the speed of cache and system memory. The most current AGP drivers for each chipset were installed for the tests.

 

AZ11/SDRAM

AZ11/VCSDRAM

AK72/SDRAM

AK72/VCSDRAM

Winbench 99

    

Business Graphics

354

353

305

301

High-End Graphics

859

853

751

745

Passmark 2.0a

    

2D Graphics

102.8

96.9

176.7

98.6

3D Graphics

162

155

173

157

3Dmark 2000

    

3D Marks

3016

3009

2827

2841

CPU 3D Marks

205

204

190

188

The Classic Athlon on the AK72 with VCSDRAM is the clear loser here, while the Tbird on the AZ11 with SDRAM is the overall performance winner in most cases – except for the Passmark graphics results. Here, the classic Athlon wins with both SDRAM and VCSDRAM, though the SDRAM scores are much better. The AK72 tests were run twice to verify the numbers. This illustrates the danger of using a single benchmark to determine the overall speed of a component.

Since the classic Athlon scored higher with both types of memory, the obvious conclusion is that the Passmark test tends to overflow the smaller cache more often, however the difference between the SDRAM and VCSDRAM scores seem to show something else. The only conclusion I can make here is that a cache miss with SDRAM causes a severe penalty, while VCSDRAM handles it much better, however the longer latency of the VCSDRAM creates a performance problem even with a lower cache miss ratio.

I/O Subsystem Performance

Once again, Winbench99 and Passmark Performance Test were used for this comparison. The most current busmaster drivers were installed for these tests. It should be noted that the hard drive used was partitioned into several logical drives. The disk benchmarks all used the D: drive rather than the C: drive. The reason for this is that the excessive I/O to the Windows swap file on C: contends with the test file I/O in the benchmarks. Alternatively, I could have put the swap file on D:, and let the benchmarks use the C: drive for their tests. Leaving the benchmark test files on C: dropped the disk drive numbers for both benchmarks almost in half.

 

AZ11/SDRAM

AZ11/VCSDRAM

AK72/SDRAM

AK72/VCSDRAM

Winbench 99

    

Business Disk

5835

5760

5820

5765

High-End Disk

12800

12800

12700

12700

Passmark 2.0a

    

Disk Mark

72.8

73.5

73.8

73.6

It is clear from these numbers that there is very little difference in disk I/O speed due to cache. This would make sense, since disk I/O uses system memory (usually via DMA) and does not rely on cache, and data read from disk is generally not going to be in cache already. It should be noted again that these numbers are averages, and that the numbers varied quite a bit even between the three runs on the exact same platform. All of the figures are within about 1.5% of each other, which is well within the margin of error of the benchmarks.

Memory Subsystem Performance

Passmark 2.0a

AZ11/SDRAM

AZ11/VCSDRAM

AK72/SDRAM

AK72/VCSDRAM

Memory Mark

108.9

112.2

109

112

Membench

AZ11/SDRAM

AZ11/VCSDRAM

AK72/SDRAM

AK72/VCSDRAM

Memory

    

Read Rate

645.16

625

571.42

571.42

Random Read Rate

186.91

181.81

185.18

185.18

Write Rate

250

289.85

277.77

322.58

Random Write Rate

57.8

62.5

62.11

68.96

Transfer Rate

129.87

156.25

124.22

140.84

L2 Cache

Read Rate

2857.14

2857.14

2000

2000

Random Read Rate

5000

5000

5000

5000

Write Rate

2222.2

2222.2

740.74

740.74

Random Write Rate

5000

5000

5000

5000

Transfer Rate

833.33

833.33

526.31

526.31

Passmark shows that these two platforms are almost identical in memory peformance, however this can only be possible if either the cache is being bypassed (unlikely) or the test intentionally overflows even a 512KB cache. In this case, we are testing system memory, which is virtually identical between these two platforms.

Membench shows some very large differences in most tests, which one would expect of a test that runs primarily from cache. The exception here is the L2 cache random read/write tests. These are suspicious, because they are identical numbers to the L1 cache test numbers (not shown here). It seems plausible that the random read/write tests in this benchmark actually were designed for smaller L1 caches, and therefore are not accurately testing the L2 cache.

It is interesting that the Membench memory write rates are faster with the classic Athlon than the Tbird. It appears that the test overflows the smaller cache, forcing it to do writes to main memory, while the 512KB cache needs to do this much less. The memory read rates seem in line with what one would expect of the smaller, but faster, L2 cache when the data does not overflow the cache, and the tranfer rate is also in line with what one would expect of VCSDRAM, since this is transferring larger blocks of data, which is one of VCSDRAMs strengths.

CPU Performance

Both Winbench99 and Passmark provide CPU specific performance results. Although Passmark breaks down the tests into add/subtract/multiply/divide for both integer and floating point operations, I decided to simply post the aggregate number here. There theoretically should not be any real difference in the actual integer or floating point units between the processors, so any difference will be the result of cache speed.

 

AZ11/SDRAM

AZ11/VCSDRAM

AK72/SDRAM

AK72/VCSDRAM

Winbench 99

    

CPU Mark 99

81.2

80.6

75.8

75.6

FPU Mark 99

4940

4940

4940

4940

Passmark 2.0a

    

Math Mark

151.2

151.1

147.7

146.2

MMX Mark

173.6

173.5

154.9

154.8

     
     
     

The results here are exactly in line with what we might expect. Both integer and floating point speed is determined by how fast the data can be fed into the processor. In this case, it is obvious that the instructions and data are primarily coming from cache, and in the case of the FPU Mark 99, directly from L1 cache. The integer performance is greatly improved by the smaller, faster L2 cache while the floating point performance is only slightly better, if at all.


Pages: « Prev   1 2 3   Next »

Be the first to discuss this article!