Benchmarks: 3D-rendering (Cinebench / Blender)
In order to measure the performance of 3D-rendering we use the officially supplied benchmarks of the open-source Blender 3D-rendering software. We also run the well-known Cinebench 15 benchmarks, both multi-threaded as well as single-threaded. Both Blender and Cinebench are known for the optimal scaling of workloads with processors that have multiple cores.
With Cinebench 15 we calculate the multi-core scaling as well. This is the multi-threaded score, divided by the single-threaded score, divided by the amount of processor cores. With optimal scaling the multi-threaded scores would be exactly the same as the amount of scores times the single-threaded score, otherwise known as 100% scaling. However, in practice this scaling is slightly lower. HyperThreading / SMT makes it so that every core can process instructions of two program threads at the same time, which means that the scaling can be over 100%.
No workload scales to multiple cores as beautifully as 3D-rendering. We see that Intel's 16-core 7960X is only a fraction faster than the Threadripper 1950X in both Blender and Cinebench. The higher IPC of the Intel architecture is compensated for by the higher clock frequencies of Threadripper. The 7980XE sets new records in both cases. With the multi-core scaling we see lower values. This can be explained by the high difference in clock frequency when only a single core is active compared with when all 16 or 18 cores are active. This is clearly done better by Threadripper.