Coretypes
A coretype is the unit of compute you pick for a Rescale job. Each coretype maps to a specific CPU or GPU architecture, a memory profile, and a storage profile, and the choice drives both runtime and cost. This page explains how coretypes are organized, how to match one to your workload, and how Coretype Sets and Coretype Collections let the platform pick on your behalf when capacity is tight.
Understanding the coretype portfolio
Rescale organizes coretypes into categories based on their primary workload fit. You browse them from the Hardware step of job setup, where each tile shows the per-core memory, network bandwidth, storage, and processor family.
The current categories are General Purpose, High-Speed Interconnect (100 Gbps and up low-latency networking for tightly coupled MPI), Large Memory (15 GB to 64 GB per core for memory-bound solvers), GPU (attached NVIDIA accelerators for AI and GPU solvers), and GPU Rendering (tuned for interactive visualization and render workloads).
Vendors span Intel Xeon (Skylake through Emerald Rapids), AMD EPYC (Rome through Turin and Genoa-X), and AWS Graviton ARM cores. Generations and vendors coexist so that you can pick the configuration that matches your solver.
Example coretypes across categories
See the full list in Coretype Explorer or on the Hardware step of any job.
Choosing a coretype for your workload
Your selection is driven by three questions:
- Is the workload compute bound or memory bound? If your solver scales linearly with cores and keeps caches hot, pick a General Purpose coretype with the highest clock speed you can afford. If the working set is large, pick a Large Memory coretype with more GB per core.
- Does the job use tightly coupled MPI? If yes, pick a coretype from the High-Speed Interconnect category. A slow interconnect dominates wall time on communication-heavy solvers.
- Does the code use a GPU solver or AI framework? If yes, see GPU and accelerators. CPU coretypes do not expose any GPU even if the solver supports one.
Rescale's published starting points:
- CFD and FEA. Emerald, Amber.
- FEA with higher memory needs. Melanite, Zinc, Topaz.
- General purpose baseline. Emerald, Ferrite, Luna, Amber.
- Large memory baseline. Zinc, Carbon, Melanite, Ruby.
For CFD, 75,000 to 100,000 cells per core is a common starting point. Run a small test at a reduced core count before you scale, since most solvers have a point where adding cores slows the job down.
Picking a core count
Core count scales with model size and time to solution, but returns diminish. Choose cores based on model size (cells, grid points, DOF), model fidelity (time accurate, precision), and whether you are optimizing for cost or for time to solution. See Hardware FAQs for starting points.
Letting Rescale pick for you
Two features let the platform choose hardware rather than pinning it yourself, and they are easy to confuse:
- Coretype Sets are an ordered list of coretype and priority pairs tried in turn, so a job starts even when your first choice has no capacity. The coretype, price and sometimes the core count can change on fallback.
- Coretype Collections are a single
Plustile mapping to several compatible coretypes, at a fixed price with your core count preserved.