[Tetralith-users] Tetralith complemented with 170 NVIDIA Tesla T4 GPUs

[Johan Raber]

Dear Tetralith User,

NSC has complemented 170 regular thin Tetralith nodes with a single NVIDIA
Tesla T4 GPU each, as well as with a faster 2 TB local scratch solid
state disk to enable new workloads to be carried out on Tetralith. A user
guide has been published at URL
https://www.nsc.liu.se/support/systems/tetralith-GPU-user-guide/

The GPU:s are of a kind well suited for development work of single or mixed
precision floating point GPU codes, including Deep Learning tasks.
TensorFlow for instance can reach performance around 2/3 that of a Tesla
V100 in mixed precision.

To fair extent, the GPUs are also suited for single-node production runs of
codes not requiring double precision, e.g. molecular dynamics codes such as
GROMACS, Amber or NAMD. For example, you can expect performance in Amber
not reachable using CPU only (of any amount), and GROMACS performs on a
Tesla T4 like two to four regular Tetralith nodes depending on load.

There is also the possibility to interactively use the GPUs on these nodes
for OpenGL visualisation much in the same way you use the GPUs for OpenGL
on the login nodes via ThinLinc. For this purpose there is now a special
version of the `interactive` tool called `interactive.vgl`, which can be
used interchangeably. For more information on how to run hardware
accelerated OpenGL at NSC, see https://www.nsc.liu.se/support/graphics/
under the heading "Running accelerated OpenGL applications". This feature
is only expected to work using ThinLinc.

Double precision codes will run on the T4s, but with *very* poor
performance. Such codes are best avoided. It is important to also know that
the primary purpose of the GPUs is to enable code development and
specialized lighter production work, they are not expected to provide a
scalable production environment where using multiple nodes and GPUs yields
significant speed-ups, if indeed any.

There are no special requirements on you to access these nodes and their
GPUs, other than adding the SLURM switch "--gres=gpu" to your SLURM
allocation command on Tetralith. For more details see the "Tetralith GPU
User Guide" referenced above. If you only need the nodes for the fast local
scratch disk (there are such use cases), you may use the SLURM allocation
switch "-C gpu" which will request a GPU-equipped node, but not its GPU if
you leave out the "--gres" switch.

Best Regards,
NSC