First cab off the rank is the Cachemem tests. This is the only synthetic test in this article. The reason for this is that synthetic tests are are only good for measuring abstract performance. What you will see in a synthetic test is either best-case or worst-case data. You won’t ever get a feel for real-world performance. Having said that, things like “processor tests ” are even worse as they are written with an architecture (or architectures) in mind. But chip designers are clever people, and love to come up with new ways of making chips faster. As soon as a new architecture (that wasn’t around when the synthetic test was written) comes along, using synthetic tests to evaluate the new architecture can be meaningless – until the test is re-written. Anyway, enough of a rant on this subject.
Cachemem tests both the latency and the bandwidth of the memory subsystem – and as this is the area where our two combatants differ – the use of such a test is valid. First up, let’s check out the available memory bandwidth.
Impressive no? Not a 100% gain, but still a fair kick in the teeth to SDRAM. Remember though – your processor will, in most circumstances, spend 95%~98% of it’s time accessing data in the Level 1 or Level 2 cache – not main memory. So here you see big differences in the numbers generated, but this performance advantage mostly won’t trickle over into huge real-life performance gains.
The other important measurement is latency. Latency is how long it takes for the chip to ask for something in main memory and actually have it ready to use. In this test, lower numbers are better.
|Score:||KT7A: 1.5||KA266: 2.0|
KT7A kicks back! Note: lower scores are better in this graph! I have expressed the memory latency time in nanoseconds, because clock cycle latency times can only be compared when the two processors are running at the same clock speed. By expressing the time in nanoseconds, we can compare latency with different speed chips. The low latency of SDRAM is something mentioned in other reviews – but not quantified. These numbers are “worst case scenarios “. The ALi chipset has probably the worst latency numbers of any of the DDR chipsets available – or so other reviews have stated. Not having any other DDR board to compare it to, all I can say for sure is that the latencies are worse than the KT133A, and once the bulk of the data spills out of any on-chip cache, significantly worse. Having said that, the latency numbers generated by the KT133A are superior to any DDR chipset equipped motherboard in production. But still, the KA266 will take a hit when main memory has to be accessed compared to the competition.
This is something that has been talked about before – even though RDRAM offered far superior bandwidth to compared to SDRAM, SDRAM systems showed as good (if not better) performance than RDRAM because the latency of RDRAM was much worse. Most applications in use today are more sensitive to latency than they are to bandwidth. One mark each. We will see if the same applies here.
|Score:||KT7A: 2.5||KA266: 2.0|
The way to do that is to see how the two compare in real-world tests. First up, Business and Content Creation Winstone 2001…
Well looky here, the IWill takes a significant lead over the Abit in the Business Winstone tests. I would not have expected this given the latency figures. But in the test where the extra memory bandwidth should have shone, Content Creation Winstone 2001, the reverse is true, the KT7A edges out the KA266. I must point out that the results obtained in Content Creation 2001 are within the margin of error. All that can be said with confidence about the Content Creation results is that the the two systems are equal performers.
What’s this about “margins of error “?
In any test, the result can only be absolute if you keep repeating the test for a large number of repetitions, and you get the exact same result. If the results vary (even a little bit), then the numbers generated have a range of values – not one absolute value. With the Winstone tests (most reviews, including this one), the Ziff-Davis recommended way to run the test is:
- Run the test five times,
- Discard the first run
- Report the highest of the next four results as the “score ” – but only if the five results are within 3% of each other. If the variation is greater than 3%, re-run the tests.
Three percent of 65 is 1.95, you can therefore only say with certainty that the score for the KT7A is somewhere between 63.05 and 65. And guess what, the score of the KA266 is within that range – which is why I claim that the two performed equally well. I have been examining this issue (and discussing it with other reviewers) and expect to come up with a better method of reporting Winstone score, but for this test, I have stuck with the Ziff-Davis recommendations.
The KA266 does however perform better than the KT7A in Business Winstone 2001 – by almost 8.6%. This is significant – and is a least a 133MHz speed grade better (i.e. moving from 1000 to 1133MHz or 1200 to 1333MHz). A point to the KA266.
|Score:||KT7A: 2.5||KA266: 3.0|
But I’m curious, why would the KA266 show such an improvement on a test that logic says shouldn’t favour it? In an earlier article at http://users.interact.net.au/~pwcl/q22001/zdrevisited.htm I exposed the weakness of the Winstone benchmarks – they are more influenced by hard disk speed than they are by memory or processor speed. Now the hard disk used in this test is exactly the same – an IBM Deskstar 75GXP – so that should not influence the result. But the hard disk controller on the motherboard is different. Lets see how VIA and ALi’s I/O performance shapes up.
A recap is in order. The VIA 686B southbridge as used on the KT7A is using the Microsoft VIA drivers with the Microsoft Pre-SP2 ATA-5 hot fix. The ALi M1535D+ southbridge is using ALi’s IDE/ATAPI driver. Anyway first up, I measured the disk access time using Winbench 99. As expected, both controller/driver combo’s returned an identical access time of 12.3 milliseconds. We can rule that out as influencing the score.
Next up, the transfer rates.
Near enough to identical. The 75GXP is getting data off of the platters and to the south bridge’s at the same rate. Strike that out as well. Next up is Business and high-end Winmarks.
Bingo! The same hard disk returns significantly different numbers using the different controller/driver combo’s. As the 75GXP is returning the same absolute physical numbers for transfer rate and access time, it is likely that the software driver is having a significant impact on the I/O performance. For some reason, the ALi driver can handle smaller, more numerous disks transactions much better than the VIA controller. Conversely, the Microsoft/VIA hard disk driver has better throughput. The scores here reflect the Business and Content Creation scores.
Business Winstone will benefit from the more efficient ALi IDE controller, as the test makes many I/O transactions with (relatively) small files. On the other hand Content Creation Winstone would prefer the faster throughput as it has less file I/O, but the files sizes are much bigger.
Those whose initial reaction to the Winstone tests would have been “here’s proof that DDR is better ” may have to think again. I will score these two rounds as one win apiece.
|Score:||KT7A: 3.5||KA266: 4.0|
On to the gaming benchmarks….
Be the first to discuss this article!