Recently, I test the DMRG and TEBD by using ITensors and ITensorGPU packages. My cpu is Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz and gpu is RTX 4070. It is found that the cpu is 5 times faster than the gpu with conserved quantum number, Did I do something wrong？
Using conserved quantities on GPU is still considered experimental, we recently got it to run but haven’t done any optimizing. The strategy we are using is quite naive but we are actively researching ways to make it faster! (EDIT: Note that you should see better speedups on GPU as you use larger bond dimensions and fewer/less sparse symmetries, with the caveat that GPUs generally have less memory available so you may run out of memory in those limits.)
Also, we recommend not using ITensorGPU.jl since it is deprecated, instead you should just use the package extensions. See Running on GPUs · ITensors.jl.
1 Like
Thank you for your reply. ：）
This topic was automatically closed 10 days after the last reply. New replies are no longer allowed.