|
Advice Beginners BIOS Guide CPUs Links Mainboards Memory Network Storage Video/Sound Cards Contact Forum SiteMap Sponsors WebNews Home |
. | . |
|
Prices: CPU Intel P4 2.4C-800 P4 2.6C-800 P4 2.8C-800 P4 3.0-800 P4 3.2-800 AMD AthlonXP XP 1700+ XP 2000+ XP 2400+ XP 2500+ XP 2700+ XP 3000+ XP 3200+ Athlon64 Athlon64 3200+ Athlon64 FX-51 Opteron Opteron 240 Opteron 242 Opteron 244 Opteron 246 Memory Corsair Crucial Kingston Mushkin OCZ |
LOSTCIRCUITS |
|
|
| VIA KT266A Chipset Bursting from the Depth of the IOQ | ||
| (Review by MS, September 3, 2001) |
Performance
The key issue of the new chipset is the increased memory bandwidth. Our theoretical calculations predicted a bandwidth increase of some 80% by moving from a burst of 4 to a burst of 8. A direct comparison between the KT266 and the KT266A should be able to shed some light on the validity of our claims from last year.
SiSoft Sandra's beta 8.21 version shows quite an increase in both ALU and FPU bandwidth, however, the difference is only in the area of some 15% over the older KT266 chipset. Keep in mind, though that SiSoft Sandra is not only measuring memory bandwidth but adds internal data transfer from the CPU cache as well as AGP transfer.
[an error occurred while processing this directive]
Above are the memory benchmark scores in SiSoft Sandra (beta 8.21) of the EPoX 8KHA as reference of the KT266 chipset, fastest setting except for tRAS set to 6T. Below are the scores of the VIA reference board to show the increase in memory bandwidth on a system level. There is certainly a difference in these benchmark results but nothing close to 80%
The pitfalls of Sandra
Very often, SiSoft Sandra is described as Stream, however, the original Stream only measures memory bandwidth along the memory bus whereas, as mentioned just above, SiSoft Sandra includes a variety of other features. For exactly this reason, the actual memory bandwidth data are diluted in SiSoft, reason enough to revert to the original Stream benchmark for a comparison of both chipsets. For accuracy, Stream needs to be run in a pure DOS environment and to accommodate the progress in technology and memory bandwidth, we increased the number of cycles 10 fold to 1000 iterations each. The results show the differences in memory bandwidth confined to the memory bus without any additional bandwidth data of the CPU and cache.
Isolating the real memory bandwidth from the rest of the system using the original stream paints quite a different picture. Listed are the different test pardigms used in StreamD. Note that the only benchmark deviateing from our predictions is copy32 which is a 32 bit copy routine that, because of it's limited size, cannot take full advantage of the increased burst length. Otherwise, the numbers are astonishingly close to those predicted in the article on [H]ardOCP one year ago. The numbers shown as Sum/5 is the average of all 5 disciplines (sum divided by 5)
The difference between the two chipsets put into percentages is almost more interesting than the pure bandwidth numbers:
Relative gain in performance of the KT266A vs. the KT266 chipset compared to the theoretical performance gain attained by increasing the burst length ffrom 4 to 8 words. Again, note that the only discipline falling out is copy32, otherwise, the predictions of our previous article are right on target.
It does appear as if the predictions from one year ago hold up extremely well, the KT266A doesn't quite reach the average 87% performance increase prediction but except for the copy32 benchmark results, the current results meet the predicted performance increase.
next page: => Gaming, Video Editing =>