Approximate cost for a trio

[gevro]

Hi, For a 30X per sample trio, what is the approximate Cloud costs of the pipeline on GoogleCloud for example? We're seeing cost estimates of ~$200 and that seems high.

[byoo]

Hi, I am another user. Out of curiosity, may I ask which version you are using? Are you running DeepVariant on a GPU?
I'm using version 1.0, and the estimates from Azure aren’t far off from yours.

[williamrowell]

Are these estimates using "standard" on-demand compute, or spot/low-priority instances? WRT workflow configuration, I'm asking whether you'd be using humanwgs_family.preemptible: true?

Generally, spot/low-priority nodes are much cheaper, but because tasks can be preempted on these nodes, the runtime is longer. My most recent experience on GCP was using Terra in 2024, but singletons typically ran ~$15-30, and trios were ~3x that price (the joint-calling steps add minimally to the cost). Runtimes were -18-36 hours, due to task preemption. Standard instances (on-demand compute) cost more (GCP docs say that spot nodes can cost up to 91% less), but can't be preempted, so they should finish faster.

I don't have a good recent estimate for Azure, but overall it's a similar issue. Azure recently discontinued their "low-priority" VMs and replaced them conceptually with "spot" VMs. Unfortunately, CromwellOnAzure was deprecated in 2025 and hasn't been updated to make use of Azure spot VMs, which means that all CromwellOnAzure tasks will now run on the standard, more expensive VMs. On Azure, this could increase the per-sample cost by ~3-4x.

[gevro]

Yes we set preemptible to true. Not sure why the charge estimates are this high.

[byoo]

@williamrowell
My estimate was based on low‑priority/spot instances, which had similar costs. In our case, networking accounts for considerable fraction of the total cost, so I assumed this could be due to our implementation. We’re planning to move away from Cromwell‑on‑Azure and focus on improving efficiency.