1{2 "_what_this_is": "Azure VM size -> EC2 instance type. Consumed by design-refs/compute.md section 5 after the rubric has chosen EC2 or an Auto Scaling group.",3 "_selection_rule": "Nearest-not-smaller on BOTH vCPU and memory within the matching family class, so a general-purpose Azure size lands on a general-purpose EC2 family rather than being re-classed. Every row records both sides' specs; a row whose aws side is smaller on either axis is a bug.",4 "_architecture": "x86_64 default. Azure VM fleets are the single biggest source of Windows workloads in this skill, and graviton.md's escape path (Windows, .NET Framework, GPU/CUDA, RDS SQL Server) fires routinely. graviton_alternative is an offer, gated on compute.md section 4.",5 "_provenance": "Authored from published Azure VM series specs and published EC2 instance specs. NOT verified against a live API.",67 "_family_classes": {8 "B": {
18 "note": "Already Arm on Azure, so Graviton is the LIKE-FOR-LIKE here, not an upgrade. compute.md section 4's x86 default does not apply to a source that is already Arm."
27 "note": "Local NVMe is the defining feature. Mapping an L-series to an m-family silently removes the local disk the workload was chosen for."
28 },
29 "M": {
30 "azure_class": "memory optimized XL",
31 "aws_family": "x2iedn / u-7i",
32 "note": "SAP-class memory. Confirm the target instance exists in the chosen region before committing; the very large sizes are not in every region."
141 "note": "GPU matches exactly; MEMORY DOES NOT - 61 GiB against 112. If the workload is host-memory bound this row under-provisions, and the GPU match is what makes it look safe."
162 "note": "AWS has no 1x A100 instance. This is a genuine architecture decision, not a size lookup: emit it as a decision for the report, never silently pick one."
181 "_gpu_caveat": "Azure and AWS do not offer the same GPU SKUs at the same granularity, and Azure sells fractional GPUs where AWS does not. Any NC/ND/NV row is a DECISION with named options and a stated trade-off, never a silent nearest-match. Route it to gpu-hpc.md when that file lands.",
182 "_unlisted_sizes": "For a size not enumerated here: read its vCPU and memory from the source, resolve its family class via _family_classes, and pick nearest-not-smaller on both axes within that AWS family. Record sizing_provenance as 'table' only if the family class was found here; otherwise 'model_prior' and say so."