メインコンテンツまでスキップ
This page is not translated yet. 日本語 is switched on so the navigation and URLs can be tested; the text below is the English original. Use the language selector in the header to switch back, or see the documentation home.

Hardware

concept · Reviewed 2026-09-14 GA

How hardware works on Rescale: you pick it per job, the platform can pick it for you, and where it physically runs is itself a choice. This page is the orientation. For coretype specifications and side-by-side comparison, see Hardware & Coretypes.

Rescale does not give you a fixed cluster. Every job names the hardware it wants, so the machine matches the run rather than the run queueing for the machine. That is the single biggest difference from an on-premises HPC environment, and most of what follows is a consequence of it.

The three things you choose​

ChoiceWhat it meansWhere
CoretypeThe instance family: CPU generation, memory per core, interconnect, GPUsCoretypes
Core count and walltimeHow wide the job runs and how long it may takeCreate and submit a job
PriorityWhether the capacity is guaranteed or interruptible, which sets the priceJob priority

Priority is the one people miss. The same coretype costs different amounts depending on whether you accept the risk of interruption, and for a checkpointing solver that risk is often worth taking. Coretype Explorer compares specifications and prices side by side.

Where hardware runs​

Compute and storage live in a cloud region, and the region determines latency to your data, which compliance regimes apply, and what capacity is available. You can let the platform place the job or pin it, which matters when a licence server or a compliance obligation ties you to one place. See Regions.

When the platform picks for you​

You do not have to choose. Two features narrow or make the decision, and admins can enforce either.

Looking for scriptable automations?

Python and Bash scripts that run around a job (for metadata extraction, notifications, and file handling) are Automations, a different capability with a confusingly similar name.

Coretype Collections​

A Collection is a single hardware tile that maps to several compatible coretypes: you pick the tile, Rescale picks the machine, and the price and your core count stay fixed. The tile name ends in Plus. See Coretype Collections for how they work and how they differ from Coretype Sets.

Compute Optimizer: machine-learning hardware recommendations​

Private Beta

Compute Optimizer is currently in Private Beta. Contact your Rescale representative to request access.

Compute Optimizer recommends a Coretype for your job based on three signals:

  • Rescale Performance Index (RPI) from software-specific ML models.
  • Rescale Maturity Index (RMI) based on capacity limits and hardware maturity.
  • Rescale Cost Index (RCI), a normalized view of prices across available hardware.

For new jobs, Compute Optimizer automatically selects a balanced Coretype based on the software you picked and any hardware policies you have set. Recommendations typically arrive in 5 to 10 seconds. If you change the core count, the recommendation can update. You can always opt out and pick hardware manually.

For cloned jobs and templates, Compute Optimizer does not overwrite the saved hardware. Instead, the recommended Coretype appears in a Recommended section and you can click to apply it.

Intelligent Resource Automation controls​

Three admin-configured features shape how the platform picks resources for your jobs. They are documented in detail under Intelligent Resource Automation; this is a quick orientation.

  • License Queuing. Holds a job until the license features and token or seat count it requires are available, rather than failing on license unavailability. License queuing is not recommended for DOE jobs and does not work for customers with a VPN but no company license proxy.
  • Region Load Balancing. Distributes compute and storage across cloud service provider regions based on capacity and region health, and respects user, workspace, or partner region preferences. If a job uses attached High-Performance Storage, region load balancing pins the job to the HPS region.
  • Coretype Sets. Defines an ordered list of Coretype and priority pairs the platform will try in sequence, helping jobs succeed when a preferred Coretype is capacity-constrained. Sets can mix ODE and ODP across providers. Billing reflects the Coretype and priority the job actually ran on.