Winstone 99
In the Winstone 99 Business benchmark, the Cool K6-III 550 took the top with an impressive score of 27.2. This is impressive because the next highest score was the PIII Xeon 1MB L2 with 24.9. The K7 550 was not far off the PIII Xeon with its 24.4, and the PIII was similarly not far off at 24.0. The Winstone 99 benchmark is a "real" test. It multitasks through each of 3 suites of applications: Microsoft, Lotus, and Corel. It doesn't simulate the tests; it actually runs through the suites.
Well, we see the K7 falling right in the middle of the PIII Xeon and the PIII. The difference between the K7 and the PIII Xeon or PIII is fairly insignificant; the .4 and .5 differences in score can be just as easily attributed to differences in the testing systems. Why, then, did the Cool K6-III beat its new big brother by a 2.8 margin?
Cache is the culprit
When multitasking, RAM makes a difference, in both speed and size. Because the system must store data from various applications and pull them up as the user chooses to use them, the system's memory affects the performance. Keeping this in mind, the K7 has a 512KB L2 cache which runs at a strange 1/3 clock speed, which calculates out to 183 MHz. On the other hand, the K6-III has a smaller 256KB L2 cache, which runs at a speedy 550 MHz. Similarly, the PIII Xeon has a 1MB cache also at full 500 MHz clock speed. The PIII has a 512KB cache at ½ clock speed, or 250 MHz.
Now knowing that the K7 and PIII both have 512KB L2 caches, one would expect the PIII to outperform the K7, but that is not the case. Referring back to the CPUMark and Disk Winmark tests that the K7 system went through, we see that in both cases the K7 scores a good margin higher than the PIII. Apparently, the architectural changes, in addition to the 183 MHz L2 cache, can equal the PIII's performance in this test, and just about put it on par with the PIII Xeon 500.
Optimization is key
What we are waiting to see, though, is optimized driver patches, and some of the other K7 versions which are scheduled to have as much as 8MB (!) of L2 cache. The speeds are supposed to increase as well, to numbers like ½ and full clock speed. When these things happen we should see a substantial speeding up of the K7. As one reader puts it:
"...points to the L2 cache being a bottleneck. Remember, the 128K L1 is split evenly between Instructions and Data. Business Winstone would be very sensitive to L2 cache speed since an empty word document is 16K and filling it up will easily overwhelm L1 cache."