Skip to content

Call for Roihu pilot projects is now open! Read the article to learn more & apply for early access.

Billing

Mahti compute billing

Jobs in Mahti's CPU partitions consume CPU BUs, either based on number of reserved nodes (node-based allocation) or number of reserved cores (core-based allocation). Memory is not billed separately, but local disk usage is billed separately. More precisely:

  • In the partitions with node-based allocation each reserved CPU node consumes 100 CPU BU per hour.
  • In the partitions with core-based allocation each core consumes 1 CPU BU per hour, and each reserved GiB of NVMe disk consumes 0.01 CPU BU per hour.

Jobs in Mahti's GPU partitions consume GPU BUs based on the number of reserved GPUs. Memory is not billed separately, but local disk usage is billed separately. More precisely:

  • Each reserved A100 GPU consumes 100 GPU BU per hour, or in total 400 GPU BU per hour for a full GPU node.
  • A slice of an A100 GPU with one seventh of an A100 (a100_1g.5gb) consumes 15 GPU BU per hour
  • Each reserved GiB of NVMe disk consumes 0.01 GPU BU per hour.

Puhti compute billing

Puhti is a heterogeneous system with CPU, GPU and IO nodes and nodes with varying amounts of memory. Moreover, it is possible to use only a fraction of a node and its various resources. The billing scheme charges Billing Units (BU) in a fair way based on the reserved amount of node resources.

On Puhti's CPU partitions the jobs consume CPU Billing Units. The CPU BU consumption rate of a compute job depends linearly on the number of requested cores, the amount of requested local storage, and the amount of requested memory. More precisely:

  • Each reserved core consumes 1 CPU BU per hour.
  • Each GiB of reserved memory consumes 0.1 CPU BU per hour.
  • Each reserved GiB of NVMe disk (if available) consumes 0.006 CPU BU per hour.
Total CPU BU = ( NCores * 1 + MemGiBs * 0.1 + NVMeGiBs * 0.006 ) * Walltime hours

On Puhti's GPU partitions the jobs consume GPU Billing Units. The GPU BU consumption rate of a compute job depends linearly on the number of requested GPUs, the number of requested cores, the amount of requested local storage, and the amount of requested memory. More precisely:

  • Each reserved GPU consumes 60 GPU BU per hour.
  • Each reserved core consumes 1 GPU BU per hour.
  • Each GiB of reserved memory consumes 0.1 GPU BU per hour.
  • Each reserved GiB of NVMe disk (if available) consumes 0.006 GPU BU per hour.
Total GPU BU = ( NCores * 1 + MemGiBs * 0.1 + NVMeGiBs * 0.006 + NGPUs * 60 ) * Walltime hours

Scratch disk billing

Puhti and Mahti have the same billing for scratch storage. Usage up to 1 TiB is free of charge.

  • Excess usage beyond 1 TiB is billed: 1 TiB consumes 5 Storage BU per hour.

ProjAppl disk billing

Puhti and Mahti have the same billing for ProjAppl storage. Usage up to 50 GiB is free of charge.

  • Excess usage beyond 50 GiB is billed: 1 TiB consumes 5 Storage BU per hour.