DeepSim is a multi-user workstation for CPU- and GPU-intensive compute jobs. Only sysadmins have sudo on this system; some users have non-root access to docker. DeepSim is currently open to a limited number of experienced users, for the purposes of testing. sysadmin notes.

DeepSim was first requested in Jun 2021, with funding agreed in January 2022, and early-access testing in April 2022. history .

Specifications

  • CPU: 64 thread; 3.5/4.2 GHz (Threadripper PRO)

  • RAM: 128GB (8 x 16GB)

  • 2 x GPU: RTX3090 24GB

  • 2TB M.2 SSD (system + home directories)

  • 4TB Seagate IronWolf Pro 3.5 (large datasets - access on request)

  • Ubunutu 20.04 (CUDA 11.4)

Performance benchmarks / early testing

  • Training ResNet50 on Imagenet: 25 minutes / epoch (both GPUs, all GPU memory used, batch = 256, tensorflow/keras). This is a 45X speedup relative to willslab_ply (batch = 16).

  • PSP exploration of EXIT model: Successfully used rocker to run these R-based simulations in a docker container.

Logging on

Request access to private github repo with log-in instructions from Andy Wills.