How much acceleration for training ?

[Jiakui]

Dear author,

In the paper, you always mentioned how much fps you achieved. Have you timed the whole training process compared to the original 3DGS ?

Thanks!

[TensorCore3DGS]

We will release the code that supports the training process soon.

However, the estimated speedup in training is not high enough (about 1.1x from DISTWAR-B) because the bottleneck of the training process is the gradient computation, the reduction process with atomic additions, and the step (update the parameters by Adam).

In our practice, we have designed methods to convert the gradient computation and reduction process into tensor operations, but it still suffers from frequent memory access, underflow(the gradients of pixels range in only about 1e-7, which underflows with FP16 ). If scaling these small gradients, catastrophic rounding errors will occur. Therefore, we are still looking for ways to accelerate the training process, particularly by reducing atomic operations and accelerating the step process.