One family.
Many kinds of power.

A dense model with P billion parameters stored at q bits per weight needs approximately P × q / 8 decimal GB for the weights alone. Add runtime data, quantization metadata, activations and the KV cache, and leave space for the operating system. The experiment uses an explicit, adjustable reserve rather than pretending that all installed RAM is free.

For a deliberately simplified, batch-one decode model, assume each generated token reads the active weights once. The bandwidth-only ceiling is bandwidth divided by weight size. This excludes compute, attention, cache traffic and framework overhead. It is a teaching upper bound, never a measured tokens-per-second result. MoE models require active-weight accounting, and batching changes reuse.

Example: 8 billion weights at 4 bits occupy 4 GB before overhead. At 100 GB/s, the idealized weight-streaming ceiling is 25 tokens/s. It is not a prediction for a real Mac.

Amdahl’s law separates a serial fraction s from parallel work. With N identical ideal workers, speedup is 1 / (s + (1 − s) / N). When 20% is serial, even infinitely many workers cannot exceed 5×. Real scheduling, unequal core types and memory contention reduce the benefit further.

This experiment uses hypothetical equal workers rather than assigning invented performance scores to Apple cores. It explains scaling; it cannot rank M1 against M6. Single-thread performance, acceleration, thermal headroom and algorithm changes are separate effects.

Example: at 20% serial work, 8 workers give 3.33× and 16 give 4×. Twice the workers produce only a 20% speedup.

Ask which device, memory configuration, power mode, application, precision, model and workload were tested. A short burst and a sustained render measure different conditions. Fanless and actively cooled systems with the same chip can diverge over time.

An Apple “up to” claim is a result under its stated test conditions, not a universal ratio. Do not chain ratios from different tests, equate TOPS at different precisions, or convert core count into a synthetic overall performance score. This atlas therefore charts specifications, not an invented benchmark ranking.

Upgrade around the bottleneck: capacity for work that cannot fit; bandwidth for streaming; GPU features for supported graphics; CPU latency for serial code; media engines for supported codecs. Check the application and the complete device before choosing.