M1 Was Designed Before ChatGPT

"Seven hundred and fifty days. That is how long the layout waited for the workload."

Two Years, Plus the Design Cycle

The M1 shipped on 2020-11-10. ChatGPT went public on 2022-11-30 — 750 days later, as the previous lesson's timeline computes. And a chip is not designed the year it ships: a processor of the M1's complexity is defined years before its release, so the decisions that gave the Mac a single memory pool were made when the largest language model in public view was a research paper. GPT-3 was described in May 2020, six months before the M1, and it was a 175-billion-parameter model that ran in a data centre; the idea that a 7-billion-parameter descendant would run on a laptop did not exist as an idea, because no such model had been released for anyone to run.

Apple's own machine-learning story in those years was on-device and small. The Neural Engine was pitched for camera processing and app-embedded models; Apple's 2022 research on running transformers on the Neural Engine targeted models that fit in a phone's memory. There is no Apple document from 2020 to 2023 that frames unified memory as a way to hold a large model. The framing arrives with MLX in December 2023, and it arrives after the community had already been doing it for nine months.

What Fit Where, When

The clearest way to see the accident is to line up memory capacities against model sizes by year. The table is public facts with public dates: consumer GPU memory, Mac unified memory, and the size of the openly available model people wanted to run.

Year	Largest consumer GPU memory	Largest Mac unified memory	Open model people wanted to run (bf16 size)	Where it fit	Evidence
2020	24 GB (RTX 3090)	16 GB (M1)	GPT-3 (not released; 175B ≈ 350 GB)	nowhere on a desk	vendor specs; OpenAI paper
2021	24 GB	64 GB (M1 Max)	—	—	vendor specs
2022	24 GB (RTX 4090)	128 GB (M1 Ultra)	—	—	vendor specs
2023	24 GB	192 GB (M2 Ultra)	LLaMA 65B ≈ 130 GB; 7B ≈ 14 GB	7B on any card at 4 bits; 65B only in a Mac's pool	Meta release; arithmetic
2024	24 GB	192 GB (M2 Ultra)	a 405B dense ≈ 810 GB; ≈ 243 GB at 4 bits	nowhere on a desk — the Mac lost this year too	vendor specs; arithmetic
2025	32 GB (RTX 5090)	512 GB (M3 Ultra)	750B-class mixtures: GLM-5.3 (753B) ≈ 1.5 TB bf16, ≈ 450 GB at 4 bits; DeepSeek-V4.1-Flash 510 GB as shipped	only a 512 GB Mac, and narrowly — the 85% thumb says no, the big-models track's measured working set says yes — or a rack	vendor specs; the household's store catalogue

Read the two memory columns. Consumer GPU memory sat at 24 GB for five years while the Mac's pool went from 16 to 512. Neither vendor was aiming at models — NVIDIA's consumer cards are gaming products, and Apple's pool grew for video and 3D — but when open models arrived in 2023 at sizes between the two, the machine whose memory had kept growing was the one they fit in. That is the mouse. It is visible in a table anyone could have drawn in 2022, and nobody did, because the row that made it interesting had not been released yet.

Code

what_fit_where.py — memory capacity vs open model size, by year·python

#!/usr/bin/env python3
"""Which desk-class machine could hold which open model, by year, at bf16 and
at 4 bits (~0.6 bytes/param). Public specs and release sizes; the point is
the shape, not the exact byte."""

GPU_GB = {2020: 24, 2021: 24, 2022: 24, 2023: 24, 2024: 24, 2025: 32}     # RTX 3090 / 4090 / 5090
MAC_GB = {2020: 16, 2021: 64, 2022: 128, 2023: 192, 2024: 192, 2025: 512}  # M1 / M1 Max / M1 Ultra / M2 Ultra / M3 Ultra
MODELS = [  # year available to run, name, parameters (B)
    (2023, "LLaMA 7B", 7), (2023, "LLaMA 65B", 65), (2024, "a 405B dense", 405), (2025, "a 753B mixture (GLM-5.3)", 753),
]

for year, name, params in MODELS:
    bf16, q4 = params * 2, params * 0.6
    gpu, mac = GPU_GB[year], MAC_GB[year]
    fit = lambda gb, cap: "fits" if gb <= cap * 0.85 else "no"
    print(f"{year} {name:16} bf16 {bf16:6.0f} GB  4-bit {q4:6.0f} GB | "
          f"GPU {gpu:3d} GB: bf16 {fit(bf16, gpu):4} 4-bit {fit(q4, gpu):4} | "
          f"Mac {mac:3d} GB: bf16 {fit(bf16, mac):4} 4-bit {fit(q4, mac):4}")

print("\nconsumer GPU memory 2020->2025: 24 -> 32 GB; Mac unified memory: 16 -> 512 GB")

Exercise

Run what_fit_where.py and add a row for the year you bought your Mac: its pool size, the largest consumer GPU of that year, and the largest open model you actually wanted to run. Which column did your model fit in, and at what precision? Then write one sentence on whether you bought the machine for that model or found the model after.

Hint

Most owners' honest answer is 'found the model after' — which is the mouse at household scale. The founder's two 512 GB machines are the exception: bought in March 2025 for exactly the row that fits nowhere else, after the mouse was already visible.

Two Years, Plus the Design Cycle

What Fit Where, When

Code

External links

Exercise

Progress

Comments 0