Nine Macs, One Household: This Quest's Laboratory

"A benchmark from a machine you cannot name is a rumour. A benchmark from a machine you can ssh into is data."

The Machines

Every measured number in this quest comes from one of nine Macs in one household, read live on 2026-09-15 over ssh. They appear here by role alias — office, server, and so on — with chip, memory and job, and never by hostname, address or owner. That is a privacy rule and a teaching rule at once: what you need to interpret a number is the silicon and the load the machine was carrying, not where it lives.

Alias	Machine	Chip	CPU	GPU cores	Memory	Role	Evidence
office	Mac Studio	M3 Ultra	24P + 8E	80	512 GB	the engines and the control planes; Pippa's home	measured 2026-09-15
server	Mac Studio	M3 Ultra	24P + 8E	80	512 GB	the inference hub: embeddings, reranking, image generation, a chat model	measured 2026-09-15
worker	Mac Studio	M2 Ultra	16P + 8E	76	192 GB	leased jobs: media scans, transcoding, transcription	measured 2026-09-15
music	Mac Studio	M2 Ultra	16P + 8E	76	192 GB	the music studio	measured 2026-09-15
macbook	MacBook Pro	M5 Max	6 super + 12 performance	40	128 GB	the main laptop	measured 2026-09-15
pro2024	MacBook Pro	M4 Max	12P + 4E	40	128 GB	laptop	measured 2026-09-15
pro2023	MacBook Pro	M3 Max	12P + 4E	40	128 GB	laptop	measured 2026-09-15
air	MacBook Air	M3	4P + 4E	10	24 GB	the electronic drum kit's Mac	measured 2026-09-15
mini	Mac mini	M4 Pro	10P + 4E	20	64 GB	the music server	measured 2026-09-15

All nine run macOS 26.6.2 and report a 16 KB page size. Every one of them ships the rdma_ctl tool, and on every one it reports disabled — which says nothing about whether the machine could use RDMA over Thunderbolt (that is a Thunderbolt 5 question, settled in the big-models track), only that nobody has turned it on.

Which Four Are the Lab, and Why

A measurement is only informative when one variable moves at a time. The plan picks four machines that make that possible:

office (M3 Ultra), pro2023 (M3 Max), air (M3) — three tiers of one microarchitecture. Same cores, same memory generation, same operating system; only the bus width, GPU count and capacity change. That is the ladder for "what does bandwidth buy".
music (M2 Ultra) — 800 GB/s against office's 819 GB/s, with the GPU one generation older. Nearly the same bandwidth, different GPU: the pair for "what does a GPU generation buy at fixed bandwidth". The other M2 Ultra, worker, would do as well, but it carries production jobs.

And the five that stay out: server is the inference hub the household's apps depend on — a benchmark that starves Pippa's retrieval is the unified-memory lesson learned the expensive way; worker is leasing jobs from the control plane; mini is streaming music; pro2024 would add an M4 point to nothing this quest measures; and macbook is an M5 Max, and the plan's rule is that the M5 generation appears as vendor claims only until it is measured deliberately, as its own evolve, rather than incidentally because it happened to be the machine the session ran on.

Read the Fleet Yourself

The code block is the script that produced the table: it runs the lesson-one card over ssh for each alias and prints one row per machine. Aliases come from your ~/.ssh/config; the script never needs to know a hostname. If you have one Mac, it prints one row — and that row is the first entry on the card you are keeping.

Code

fleet_table.py — one row per Mac, over ssh, by role alias·python

#!/usr/bin/env python3
"""Print a fleet table from ssh aliases. Aliases resolve in ~/.ssh/config;
this script never sees a hostname, and neither should a quest."""
import subprocess
import sys

ALIASES = sys.argv[1:] or ["office", "server", "worker", "music", "macbook",
                           "pro2024", "pro2023", "air", "mini"]

# One line per field. The cluster line reads hw.perflevelN.name for every level that
# exists, because the names move with the generation (M3: Performance + Efficiency;
# M5 Max: Super + Performance) -- a fixed P/E pair would mislabel the newest Mac.
REMOTE = (
    "sysctl -n machdep.cpu.brand_string; "
    "c=''; for n in 0 1 2 3; do name=$(sysctl -n hw.perflevel$n.name 2>/dev/null); "
    "[ -z \"$name\" ] && break; c=\"$c${c:+ + }$(sysctl -n hw.perflevel$n.physicalcpu) $name\"; done; echo \"$c\"; "
    "sysctl -n hw.memsize hw.pagesize; sw_vers -productVersion; "
    "system_profiler SPDisplaysDataType | awk -F': ' '/Total Number of Cores/ {print $2; exit}'"
)

print(f"{'alias':8} {'chip':10} {'CPU clusters':28} {'GPU':>4} {'GB':>4} {'page':>6}  macOS")
for alias in ALIASES:
    try:
        out = subprocess.run(["ssh", "-o", "BatchMode=yes", "-o", "ConnectTimeout=5", alias, REMOTE],
                             capture_output=True, text=True, timeout=30).stdout.split("\n")
        chip, clusters, mem, page, macos, gpu = (out + ["?"] * 6)[:6]
        print(f"{alias:8} {chip.replace('Apple ', ''):10} {clusters:28} {gpu:>4} "
              f"{int(mem) // 2**30 if mem.isdigit() else '?':>4} {page:>6}  {macos}")
    except (subprocess.TimeoutExpired, OSError) as exc:
        print(f"{alias:8} unreachable ({exc.__class__.__name__})")

Is RDMA over Thunderbolt enabled here? (it is a Recovery-mode switch)·bash

rdma_ctl status
# disabled        <- every Mac in this fleet, 2026-09-15
# The tool ships on every Apple silicon Mac, Thunderbolt 4 ones included.
# Its presence says nothing about eligibility; see the big-models track.

Exercise

Open your Mac card for real now: a text file with one row per Mac you can reach, in the columns of the table above plus a "job" column in your own words. Then mark which rows could form a one-variable pair or ladder — same generation across tiers, or same bandwidth across generations — and which could not, and why. If you have exactly one Mac, write down which single published spec you would need a second machine to isolate.

Hint

Two laptops of different generations at the same tier (an M3 Max and an M4 Max, say) change memory technology AND GPU generation at once — an honest pair only if you accept that you cannot tell the two apart. The same tier in the same generation with different memory sizes is a capacity-only pair, and the memory track will want one.