TX1 / TX2 Performance Comparison


Jetson TX1 and TX2 has same CPU spec.

  • Jetson TX1 CPU : Quad ARM A57/2 MB L2
  • Jetson TX2 CPU : Quad ARM A57/2 MB L2 + HMP Dual Denver 2/2 MB L2

I wonder if computational performance of a single core.
Does TX2 have better CPU performance than TX1?

I can’t answer based directly on CPU spec, but I believe caching may be used differently (better) in the TX2 (I couldn’t tell you exactly how, but there were places in the TX1 where caching was not enabled and the TX2 solved issues and is able to use caching in those corner cases). If I am right about this (and I am not positive), then there will be better performance on the TX2 in some niche cases even if the cores are exact matches.

If you enable the two Denver cores, then yes, the TX2 has 50% more cores than the TX1, and thus can compute 50% more on the CPU, assuming your workload can be parallelized.

The standard TX1 kernel and userland was 32-bit, and the standard for TX2 is 64-bit; presumably because the TX2 has 8 GB of RAM.

If you are massively memory bound and your data structure has lots of size_t/long/pointers in it, then the 64-bit mode will be slower, but you can compensate for that by building and running 32-bit binaries with a multiarch install.

If you set the clock speeds differently, then the core performance will vary accordingly. (Make sure to set the TX2 to nvpmodel -m 0 and run jetson_clocks.sh to turn on the highest performance.)