| « | » |
From now on we completely switch to CPU RM 2003, as it's quite stable now. "Beta" status in this case means that our programmers have not yet added all the functions planned in the release. Note that the situation with performance has slightly changed - processors on NetBurst core demonstrate higher results now. But having compared the absolute values with the results of the previous tests, we didn't notice CPUs (including AMD) operating any slower, it's just that the P4 code has been optimized. And certainly, the improved SSE3 support has its effect in case of Prescott. Surely AMD processors have their own features to be supported (e.g. AMD64 architecture), however the protracted release of 64-bit Windows for AMD64 does not allow to work on this problem.
Considerable breakaway of Prescott is the only issue worthy of a discussion here. It cannot be explained solely by the increased frequency: it is higher than in Pentium 4 XE only by 8%, while according to the test results Prescott @ 1066 FSB outscored it approximately by 16%. We are only to assume that those additional 8% of gain are due to the SSE3 optimization. However, we should also note the drastically improved hardware prefetch in Prescott core - it should have made its presence felt with fast memory.
In case of Intel processors, the renderer operated in two-threaded mode (as they both support Hyper-Threading), and basically brought no surprises - NetBurst core demonstrated the best results here, as it did before.
We returned (or, one might say, tried to return) to CPU tests using SPEC viewperf, as the 8th version of this benchmark is considerably updated and contains a lot of potentially nice applications: CATIA, Maya, SolidWorks, Pro/ENGINEER. But we must make one important comment here: viewperf cannot be considered a test, which results fully reflect processor performance in the given applications, because it reflects only one aspect - rendering speed. Besides, viewperf results depend on a video card and its drivers to a considerable degree.
The highest rendering speed in 3ds max was demonstrated by Intel processors, but this is actually not that important (remember that the AMD platform acts as a guest star in this test and it cannot be directly compared with systems based on Intel CPU, not being a top solution). Interestingly, both Pentium 4 processors are on a par. Most likely, this subtest is limited by the performance of a video card.
The situation here resembles that of the previous test.
...The same picture again...
And here we can speak of the obvious heavy load on the processor and some architectural preferences. The latter are most likely expressed by the performance - pipe length relationship - the results are arranged in neat regular steps, the best result being demonstrated by Athlon 64, and the worst - by Prescott.
And the Maya subtest is evidently cache-hungry. Of course, you cannot draw conclusions only from a single test of three processors, but Prescott hardly managed to catch up with Athlon 64 3500+, having a cache twice as large and a huge frequency advantage. But P4 Extreme Edition outscored them all anyway, most likely due to its 2 MB L3.
...Video card is the limit here again... To all appearances, if the AMD processor had a higher clock and a larger cache, all the three columns would have been aligned in a row.
Approximately the same situation, only the gap is still smaller.
Actually, there is nothing to comment on.
Well, "return to SPEC veiwperf", despite its updated set of applications, brought no special surprises. To all appearances, we shall publish only the general mark for the entire test in the next articles: the majority of separate subtests are hardly worthy of rapt reviews.
On the whole, 3ds max 6 performance tests with the help of a script from www.spec.org should be acknowledged more informative than the set of scenes we used for rendering. So we decided to leave the above mentioned set of scenes for the expanded tests and use the package from SPEC in usual tests. The fact is that it reflects better the work process in 3ds max, because it combines the rendering process proper with other operations. You can even see how it happens by downloading a 10 MB video file in Windows Media Video format (certainly, it does not show the entire testing process, but you can get an idea). We provide the overall time of script execution as a resulting value, though more detailed statistics are available as well.
The "future" Prescott demonstrates the best result, though it can be written off due to higher frequency (the gain is actually comparable). The system based on the new ATI chipset and Athlon 64, unfortunately, didn't pass this test (it would constantly freeze somewhere in the middle). Nothing serious, we'll write it off for a "raw" sample.
The new (for our method) Adobe Photoshop version and a new test script. It was considerably expanded and is still in the development stage (basically, the number of various operations is changed to obtain better balance), so these results must be considered interim. The script now uses the following set of actions and filters:
High memory operating speed, "optimization" of major used functions for multi-threaded operation (this was not an end in itself, they are actually used more often than the other!), relatively large L2 cache, and frequency of course - that's the list of key parameters that we think procured victory for Prescott. And even the huge L3 did not save P4 XE from defeat in this test...
| « | » |