The Big-Memory Max Was Pitched at Video

"Read the press release for the 64 GB laptop. Every sentence about memory is about frames, not tokens."

October 2021, in Apple's Words

The M1 Pro and M1 Max were the first Apple silicon chips whose memory and bandwidth numbers would later matter to inference: 400 GB/s and 64 GB on a laptop, in 2021, when the biggest consumer graphics card had 24. Here is what Apple said the numbers were for, quoted from the release of 2021-10-18:

"M1 Max delivers up to 400GB/s of memory bandwidth — 2x that of M1 Pro and nearly 6x that of M1 — and support for up to 64GB of unified memory."
"M1 Pro and M1 Max also feature enhanced media engines with dedicated ProRes accelerators specifically for pro video processing."
"M1 Max transforms graphics-intensive workflows, including up to 13x faster complex timeline rendering in Final Cut Pro compared to the previous-generation 13-inch MacBook Pro."
"…enabling creatives like 3D artists and game developers to do more on the go than ever before."
"…allowing playback of multiple streams of high-quality 4K and 8K ProRes video while using very little power."

And the release's only machine-learning sentence: "A 16-core Neural Engine for on-device machine learning acceleration and improved camera performance." The memory was for editors. The bandwidth was for 8K streams and 3D scenes. The Neural Engine was for the camera. Thirteen months later ChatGPT would give those same two numbers a meaning nobody in the release had in mind.

Why the Video Pitch Produced the Right Shape

This is the part of the mouse story that is not luck in the sense of coincidence; it is luck in the sense of a right answer arrived at for a different question. Professional video is also a bandwidth-and-capacity workload: an 8K ProRes stream is gigabytes per second, a colour-graded timeline with effects holds many frames resident, and a 3D scene wants its textures in memory. The engineering target — feed a wide GPU from a big, fast, shared pool without copies — is the same target a language model's decode phase would set two years later. Apple built the shape for frames, and frames and tokens turned out to want the same shape.

The clearest sign that the pitch was sincere is the media engine. ProRes accelerators — fixed-function silicon for one video codec — are an expensive commitment to a specific customer, and no one builds them as a side effect of an AI plan. The CPU track showed that this household's Ultra carries four of them. They are the fossil record of what the big-memory Max was for.

What Changes When You Read It This Way

Two things. First, the M1 Max's bandwidth is not "an AI feature" and never was: it is a video feature that AI inherited, which is why Apple's 2021 marketing has no inference claims to be held to and why this quest's vendor-claim callouts start in 2025. Second, the "admit what must be admitted" clause has a precise object: not that Apple foresaw the workload, but that Apple built the right shape for a different workload and then, once the mouse was visible, spent real silicon following it. The next three lessons are that follow-through, and what it did not fix.

Code

frames_vs_tokens.py — two workloads, one shape: bytes per second and bytes resident·python

#!/usr/bin/env python3
"""An 8K ProRes stream and a 27B model's decode want the same two things from
memory: sustained bandwidth and a large resident set. Physics only."""

# Apple's ProRes white paper: 3840x2160 at 24p, ProRes 422 HQ = 707 Mb/s; 8K is 4x the pixels, 60p is 2.5x the frames
prores_4k_mbps = 707
prores_8k_gbs = prores_4k_mbps * 4 * 2.5 / 8 / 1000        # GB/s per 8K60 stream, compressed; decoded frames are far larger
frame_8k_bytes = 7680 * 4320 * 2 * 3                         # 16-bit RGB, one decoded frame
streams = 4

model_gb = 14.42                                              # Qwen3.5-27B 4-bit, bytes per token
tokens_per_s = 32.6                                           # measured, office

print(f"video: {streams} × 8K ProRes streams ≈ {streams*prores_8k_gbs:.1f} GB/s compressed in; "
      f"one decoded 8K frame = {frame_8k_bytes/1e6:.0f} MB; a 2-second scrub buffer at 30 fps ≈ {frame_8k_bytes*60/1e9:.1f} GB resident")
print(f"tokens: 27B decode at {tokens_per_s} tok/s reads {model_gb*tokens_per_s:.0f} GB/s; {model_gb:.1f} GB resident, always")
print("\nboth: a wide bus, a big pool, no copies. Apple built it for the first line; the second arrived later.")

Exercise

Run frames_vs_tokens.py. Then open the 2021 release yourself and count: how many sentences mention memory or bandwidth, how many of those are about video, 3D or rendering, and how many are about machine learning? Write the three counts on your card and one sentence on what a reader in October 2021 would reasonably have concluded the 64 GB was for.

Hint

The count is lopsided enough that the sentence writes itself. The interesting part is the last one: the reader would have been right, and would have been right for two more years.