Coretype Explorer

how-to · Reviewed 2026-11-03 beta

Coretype Explorer plots Rescale coretypes against each other on the specifications that decide a hardware choice: memory per core, clock speed, networking and storage. Use it to narrow a shortlist on paper, then benchmark the survivors on a real workload.

Open beta

Coretype Explorer is in open beta and is not enabled by default. Contact Rescale Support to have it turned on for your workspace.

Who it is for

Administrators use it to understand what hardware their organization exposes and to make recommendations, or to fix a sensible default into a job template. Engineers use it when tuning a workload for cost, time to solution or memory headroom, and can open it inline while picking hardware for a single run.

Open Coretype Explorer

Where you open it from depends on what you are doing: inline from a run you are configuring, or from an admin portal when you are surveying what the workspace offers rather than choosing for one job.

From	How
Setting up a job or workstation	Where you select a coretype, click Compare coretypes using the Coretype Explorer
Either admin portal	Choose Coretype Explorer from the Tools menu. It is in both Organization Admin and Workspace Admin

Compare on the graph

The Explorer opens on an interactive graph rather than a table.

Down the left you choose which coretypes to compare, and you can filter by category first to cut 200-plus coretypes down to the handful worth looking at. The categories are the same ones described under coretypes: General Purpose, High-Speed Interconnect, Large Memory, GPU and GPU Rendering. Filtering to High-Speed Interconnect before comparing is the quickest way to shortlist for a tightly coupled MPI job.

On the right you choose which metric to plot and switch the graph style between line, radar and bar. Radar is the useful one when you are weighing several specifications at once and want to see the shape of a trade-off rather than a single ranking.

Inspect the full specifications

Scroll to the bottom of the coretype list and click Inspect for the complete specification table. The fields available are:

Field	What it tells you
Coretype	The coretype name, for example Emerald, Kyanite or Albite
Processor Info	Vendor, family and generation of the underlying CPU
Cores / Node	How many cores one node provides
Base Clock Speed (GHz)	CPU base frequency, a reasonable proxy for per-core throughput
Network Bandwidth (Gbps)	Inter-node bandwidth. High values signal a tightly coupled fit
Memory / Node	Total RAM per node
Memory / Core	RAM per core, the number that decides whether a memory-bound solver fits
GPU Info	NVIDIA model and memory, where a GPU is attached
GPUs / Node	How many GPUs one node provides
Storage / Node	Local NVMe or SSD per node
Storage / Core	Local storage per core
Price (ODE) and Price (ODP)	Price at each of the two on-demand priorities. See priority models

Two further fields, Instruction Set and Network Latency (μs), appear in the list but carry no data: the coretypes API returns an empty value for both on every one of the 80 coretypes sampled in November 2026. Treat them as placeholders rather than as something to sort on, and use Network Bandwidth when you need a proxy for interconnect quality.

Narrow a shortlist

1. Start from the workload's binding constraint. For CFD that is usually network bandwidth and clock speed; for FEA with large matrices it is memory per core; for AI training it is GPU model and GPU memory.
2. Filter by the matching category, then plot that metric across the survivors.
3. Bring price in second, not first. A coretype that cannot hold the working set is not cheaper, it just fails later.
4. Check storage per core if the solver writes large checkpoints or scratch files locally.
5. Benchmark what is left. Specifications narrow the field; they do not settle it. Run the real workload across the finalists in a performance profile.

Coretype Explorer is also worth opening before you pin members of a coretype set, since a set whose members differ sharply in cores per node or memory per core will behave inconsistently when it falls back.

Related

• Coretypes — what each coretype is built for, and the category definitions
• Performance Profiles — benchmarking a real workload across the shortlist
• GPU and accelerators — choosing and using GPU coretypes
• Priority models — what ODP, ODR and ODE mean
• Hardware FAQs — common questions about hardware selection