Table 2 Average computing time each algorithm takes to process a single particle (in milliseconds).

From: Artificial intelligence for improved fitting of trajectories of elementary particles in dense materials immersed in a magnetic field

Processor

Parallelisation

SIR-PF

RNN

Transformer

CPU

single-thread

435.71 ± 5.18

91.16 ± 1.17

12.25 ± 0.19

 

multi-thread

-

82.22 ± 1.00

6.58 ± 0.04

 

batch_size = 1

-

31.27 ± 0.99

8.96 ± 0.31

GPU

batch_size = 16

-

4.02 ± 0.12

1.24 ± 0.12

 

batch_size = 64

-

1.43 ± 0.05

0.71 ± 0.04

  1. The test shows the average results of running the three methods (sequential importance resampling particle filter (SIR-PF) with all hits, recurrent neural network (RNN), and transformer) on the same ten random subsets of the testing dataset consisting of 10,000 particles each. CPU: AMD EPYC 7742 64-Core 3200 MHz Processor, GPU: NVIDIA A100 Tensor Core (8 GB of memory). Note that the SIR-PF implemented does not support multi-threading nor GPU computation since it is out of the scope of the article; parallelising the computation for the RNN and Transformer becomes trivial, thanks to PyTorch. The parameter 'batch_size' indicates the number of particles processed together in each step.
Back to article page