Latch supports using Nextflow accelerators to allow processes to leverage GPU compute.

​
Caveats

GPU enabled nodes have very strictly constrained compute parameters. As a result, specifying an accelerator will override any cpu/memory settings, with overridden values changing depending on the chosen accelerator.

​
Supported Accelerators

Latch supports the following accelerator types:
  1. nvidia-t4: This allows a process to use an Nvidia T4 GPU for compute. A process specifying type: 'nvidia-t4' can use only 1 GPU. Choosing nvidia-t4 will override and set cpu: 7 and memory: "30Gi".
  2. nvidia-a10g: This allows a process to use an Nvidia A10G GPU for compute. A process specifying type: 'nvidia-a10g' can use only 1 GPU. Choosing nvidia-a10g will override and set cpu: 64 and memory: "256Gi".
  3. nvidia-v100: This allows a process to use an Nvidia V100 GPU for compute. A process specifying type: 'nvidia-v100' can choose between 1, 4, or 8 GPUs. Choosing nvidia-v100 will override and set resources differently depending on the number of GPUs specified:
    • Choosing 1 GPU will override and set cpu: 7 and memory: "48Gi",
    • Choosing 4 GPUs will override and set cpu: 30 and memory: "230Gi", and
    • Choosing 8 GPUs will override and set cpu: 62 and memory: "400Gi".