CineBench is one of the tests that measures processor performance in rendering. So it is a real load that is a reflection of the performance of the real application. However, while about 6 years ago (which is not that long ago) standard desktop processors offered a maximum of 8 threads (typically four cores) and servers a maximum of 44 threads (22 cores), today desktop processors with 32 threads (16 cores) are commonly available. and server processors with 192 threads (96 cores), and there are also products with up to 256 threads (128 cores) behind the door.
It went fast. Too fast for some software companies, because they didn’t adapt their software as fast as needed. It is CineBench R23 that is mentioned from time to time in the middle of the speech, that it handles a maximum of 256 threads. However, users point out that compared to some other tests that can use a higher number of threads, CineBench R23 stops scaling optimally already for hardware configurations with a lower number of threads than 256.
So how is it? Empirically, it is easy to see that CineBench renders a tiled image, where each thread is assigned one tile. If the processor supports 8 threads, eight tiles are rendered, when the first one is rendered, the next one is assigned to that processor thread. Therefore, it is not possible to use more threads at once than the number of tiles on which the image is spread. One might come across the opinion that when CineBench R23 splits the image into 220 tiles by default, no more than 220 processor threads can be used. The reasoning is correct, the problem is that on processors with higher thread counts, CineBench reduces the tile size, so instead of 11×20 (220) tiles, the image can be spread over 88×107 (9416) tiles.
However, that may be the problem. Changing the size of the tile changes the nature of the load. With a smaller tile, the demands on the cache capacity decrease and rather the demands on its speed increase. Thus, the 8-core and 128-core processors in CineBench are not served identical workloads, which may result in slightly different than expected results for processors with high core counts.
The combination of the facts described above well explains several phenomena that we may have encountered recently. For one thing, it’s the reason why 128 cores Zen 5 achieves higher performance than 192 cores in CineBench R23 Zen 4 (192-core configuration Zen 4 is used only ~67%, while 128-core Zen 5 completely). But it also explains the big halo around the world record in CineBench R23 on an overclocked and nitrogen-cooled Xeon Sapphire Rapids. Epic Zen 4 they are more powerful, but CineBench R23 does not use all their threads, so the processor runs “on three cylinders”.
It is therefore necessary to very carefully evaluate any leaks about the performance of the upcoming processors measured precisely on CineBench R23. A number of resources are unaware of these limits, so we are concerned with the 128-core configuration test Zen 5 they could be met with the opinion that actually Zen 5 achieves 50% more performance per core than Zen 4 and this is only a sample, so the final performance will certainly be much higher (!!!). Not really. The only problem is that the 192-core (384-thread) configuration scores Zen 4 is actually a score of 256 threads, so the performance per thread actually for the sample tested Zen 5 it reaches only 10-15% higher value than the serially produced one Zen 4. In short, as can be expected for a sample in condition > a year before release.
