li012589 / trgbenchmark

Using simple TRG code to benchmark different tensor implements.

Benchmark of Different Tenosr Implements

Currently, this test focus on three different implements: using pytorch; using julia and using itensor.

The result is

Implement	Time(average over 10 runs)
pytorch	1.73s
julia	9.31s
itensor	9.75s

The scaling curve with iterations is

The scaling curve with maximum cuts is

Details

This test is done on a mid 2014 MacBook Pro with a 2.2GHz Intel i7 processor and 16G of memory. And the pytorch test doesn't utilize a GPU.

Library versions:

name	version
pytorch	1.0.0.dev20180921(nightly)
julia	1.01
itensor	2.1.1

To perform this test on your machine, run

python ./test.py
python ./testscaling.py

About TRG Algorithm

A beautiful tutorial of TRG by itensor

trgbenchmark's Issues

Some fix on presentation of benchmark

• Table for Time(average over 10 runs): Show mean +/- error ? Can have less digits
• maximumcut.png use log plot ? (since it is expected to have power law scaling)
• add links to the original itensor page