What speedup can be expected from GPUs?

[ajthomas1949]

• Orion version: 0.3.3
• Python version: 3.6.15
• Operating System: Linux 4.14.287-215.504.amzn2.x86_64 x86_64

Description

hi, I wanted to compare training and anomaly detection times using TadGAN on CPus and GPUs.

What I Did

We created a docker version on an AWS p2.xlarge EC2 instance with 1 NVIDIA K80 GPU. Training and test data
sets had about 9000 data points each. Using 5 epochs, I saw that the training and anomaly detection times 
using the 'tadgan_gpu' pipeline were about 2x faster than usng the regular 'tadgan' pipeline. Is this in line
with experience of others? I was hoping for a greater speed-up!

[sarahmish]

Hi @ajthomas1949!

From my experience, the GPU counterpart should be faster but this depends also on your hardware. Here's a little experiment I just ran:

Intel(R) Core(TM) i9-7920X CPU @ 2.90GHz
NVIDIA TITAN RTX 24 GB

epochs	CPU	GPU	speedup
5	104.2	16.2	~6.5x
10	179.2	30.7	~5.8x
20	344.8	60.5	~5.6x