Pick a source instance type.
The finder ranks the closest type on every other cloud.
Pick an instance type on one cloud — the finder translates it to the closest match on every other cloud. Each result carries a transparent match score, so you can see exactly why it ranked where it did.
Three additive axes — vCPU, memory, category — sum to a 100-point base. That
base is then multiplied by an architecture factor: ×1.0 when the architecture
matches, ×0.55 when it doesn't. Architecture is a multiplier rather than just
another delta because an image built for x86 will not boot on Arm — a cross-architecture
“equivalent” is capped at 55 however well its sizes line up. Sizing
axes use a log₂ ratio penalty, so a doubling matters the same whether you go
from 2 → 4 vCPU or 64 → 128. Nothing here is a black box: the chips under
every result spell out the exact deltas that produced its score. Price and region counts
shown on results come straight from each instrument’s live data.json and
never affect the score — they are on-demand list prices (hourly, per the source
cloud’s billing currency), not what a committed or spot workload would pay.
The dominant axis. An exact vCPU match keeps all 50 points; every doubling or halving away from the source costs the full weight.
50 × max(0, 1 − |log₂(cand ÷ src)|)
Same log-ratio shape as vCPU. If either side's memory is unknown, this axis is dropped and the base is renormalised across the remaining axes — the result is still scored, just on fewer axes.
38 × max(0, 1 − |log₂(cand ÷ src)|)
General, compute, memory, storage, accelerated, HPC. An exact category earns 12; an adjacent one — e.g. general next to compute or memory — earns 6; anything else earns 0.
exact → 12 · adjacent → 6 · else → 0
x86 vs Arm — applied as a multiplier on the 100-point base, not an additive axis. Same architecture leaves the score untouched; a mismatch scales it to 0.55 and is flagged red.
same arch → ×1.0 · differs → ×0.55
Per cloud, the finder shows the three highest-scoring types. The top score colours by confidence band, and the chips below each row break the score back down into its parts.
Google Cloud's public dataset publishes machine types but not memory, so its vCPU count is
read straight from the documented type-name suffix (n4-standard-8 = 8 vCPU) and
its memory is left unscored. Some Azure burstable sizes and OCI / OVH bare-metal and storage
SKUs ship without published vCPU or memory — those are excluded as candidates rather
than guessed at. Specs come verbatim from each instrument's own dataset; none are invented
here.