As you probably know, we’re finishing the Arion v2.5.0 core. Because of this, we have compiled and packed a new version of ArionBench, which is available for download on our website.
The time scale in the new ArionBench has been adjusted so its results match those of previous versions, despite the core is a bit faster now.
Besides that, the bench offers a new feature. Now it is possible to specify a custom number of passes to render for. Some of our customers are evaluating machines so powerful that the bench takes too little to render, producing benchmark measurements that are not stable enough. If that is your case, you can invoke AB this way:
arionbench_64 -hardware -passes:10000
This will force the bench to render for 10000 passes (the default is about 1500). If you own a ‘normal’ machine, you can simply use one of the .bat files included in the .zip package, like before. The bench has been designed to take about one or two minutes to complete in a more or less powerful machine. But, as I explained, in some cases that may not be enough (in configurations with 8 TITANs, for example).
New ArionBench v2.5.0
We have been running ArionBench in some very fat machines lately and we have found out that in some cases (when there are many GPUs) it seems that the benchmark value gets a little better if a higher number of passes is used. In some cases the difference between using the default benchmark (1500 passes) or a much higher number of passes (e.g., 10000) can be significant (+30%). The reason for this can be summarized in the following sentence: “it may take a little while for all the GPUs to be fully loaded and working at full capacity”.
For this reason, it is recommended that if you want to benchmark your machine as accurately as possible, you must set a number of passes that is long enough. This is specially important in very powerful machines with many GPUs (where the benchmark takes less than one minute to compute), and not so important in ‘normal’ machines where the benchmark takes 2-3 minutes or more to complete.
Note that (at least for now) we have decided to fixate the number of passes to a number that will never render for an oddly long amount of time, despite this fixed number may be too short in some configurations.
The top entries in the ArionBench page are now a bit outdated, since the v2.5 benchmark (with a larger number of passes) will produce an even higher output. If you submitted one of those entries yourself using v2.4.x or older, please feel free to enter a new, updated, v2.5 entry.
Thanks for watching!
UPDATE: We added a page about ArionBench in the Knowledge Base: