New GPU nodes for interactive use on Anansi#

../../../_images/zen5-closeup-01.jpg

We are happy to announce the immediate availability of 8 new GPUs on Anansi, our cluster for interactive and visualization workloads directly connected to VUB’s OnDemand web portal.

These 8 new GPUs are NVIDIA L40S models with 48 GB of video memory and 91.6 TFLOPS of FP32 compute power. They are very powerful GPUs than can outperform the A100 in Hydra on certain tasks, such as AI inference and graphic processing. However, more intensive tasks such as AI training or high-precision simulations (e.g. molecular dynamics) are better suited to the A100s.

The CPUs paired with the L40S are AMD EPYC Zen 5 with 64 cores per node (16 cores per GPU) and 6 GB of system memory per core (96 GB per GPU). The systems with L40S are connected to the high-speed Infiniband network at 200 gbps. Both the CPU architecture and network connectivity are the same as the zen5_mpi partition in Hydra, facilitating using these new compute nodes for testing and debugging of jobs targeting that partition.

You will find these new GPUs in the ada_gpu partition of the Anansi cluster. As with the rest of Anansi, these systems are configured for shared usage in interactive and light workloads. Hence, each GPU can be used by up to 4 jobs (or OnDemand sessions) at the same time for a maximum time of 12 hours.

Furthermore, we made the following changes in our OnDemand web portal:

  • we added a new application with an improved graphical desktop environment, with lower latency and seamless copy/paste functionality on Chrome based browsers. You can find it in the testing category as Desktop v2.

  • the form to select GPU resources now uses fractions of a GPU. The use of GPU fractions combined with a maximum of 1 GPU device per job (100% GPU fraction) will increase the availability of these resources and minimize queue time.

Contact us at VUB-HPC Support if you have any comments or questions about these changes.