Performance Comparisons

[PrinceSajjadHussain]

How does Burn compare to other frameworks in your experience?

[Thor-x86]

Other frameworks? My old laptop can only use CPU on PyTorch and it takes around 30 minutes for 50 epochs of LSTM with 1000 rows of time-series data.

Meanwhile, burn with ndarray backend can manage to achieve 1 minute on modern-lstm example. That's only with a single core (I have no idea how to enable multi core).

However, when I enabled Vulkan, it became 4x slower than ndarray (4 minutes). Then, I checked amdgpu_top command:

From there, I noticed that the CPU load is 138% whereas GPU only 66%. I suspect there is either too many .clone() or CPU-GPU data transaction. I'm wondering if we can compile SPIR-V ahead of time and reuse that for entire training process.