How Memory Bandwidth Has Changed Across Apple Silicon Generations
10 min read · updated August 11, 2026
Apple publishes a memory bandwidth figure for almost every M-series chip, and the sequence is not the smooth doubling the marketing arc implies. It has gone backwards once at the Pro tier, plateaued at the Ultra tier for three generations, and jumped sharply at the Pro tier once. Each of those is a direct multiplier on local decode speed.
The published figures
Every figure below was read from an Apple Newsroom announcement or an Apple technical specification page on 11 August 2026, and each is Apple’s own number rather than a measurement.
generation base Pro Max Ultra M1 not published 200 GB/s 400 GB/s 800 GB/s M2 100 GB/s 200 GB/s 400 GB/s 800 GB/s M3 100 GB/s 150 GB/s 300 or 400 GB/s 819 GB/s (M3 Ultra) M4 120 GB/s 273 GB/s 410 or 546 GB/s — M5 153 GB/s 307 GB/s 614 GB/s —
Sources, in the order the rows appear: Apple’s M1 Pro and M1 Max announcement for 200 and 400 GB/s; the M2 Pro and M2 Max announcement for 200 and 400 GB/s and for the statement that M2 Pro’s figure is twice that of M2; Apple’s MacBook Pro (14-inch, M3 Pro or M3 Max) technical specifications for M3 Pro at 150 GB/s and M3 Max at 300 GB/s in the 14-core-CPU configuration and 400 GB/s in the 16-core; the Mac Studio specifications for M4 Max at 410 and 546 GB/s and M3 Ultra at 819 GB/s; the Mac mini specifications for M4 at 120 GB/s and M4 Pro at 273 GB/s; and the M5 Pro and M5 Max announcement for 307 and 614 GB/s, with the M5 announcement giving 153 GB/s for the base chip.
The number Apple never published
The original M1 is the one gap, and it is worth being explicit about rather than filling in from a forum. Apple has never put a GB/s figure for the base M1 in a Newsroom post or on a technical specification page. What it published were two ratios, six years apart: the M1 Max announcement describes 400 GB/s as nearly six times that of M1, and the M5 announcement describes 153 GB/s as more than twice that of M1.
from "nearly 6x": M1 is a little above 400 / 6 = 66.7 GB/s from "more than 2x": M1 is a little below 153 / 2 = 76.5 GB/s
The two constraints bracket the base M1 somewhere in the high sixties, and the figure widely repeated elsewhere sits inside that bracket. This page does not assert it, because Apple did not, and a derived range with its reasoning shown is worth more than a precise-looking number with no source behind it.
Where the line went backwards
The M3 Pro is the anomaly in the table: 150 GB/s where the M2 Pro was 200, a twenty-five per cent reduction between consecutive generations of the same tier. Apple did not present it as a reduction and the chip was faster in most respects, but for local inference — where bandwidth is the whole constraint at batch size one — an M3 Pro has a lower ceiling than the M2 Pro it replaced.
The M4 Pro then went the other way and further than a recovery: 273 GB/s, which Apple’s announcement describes as a seventy-five per cent increase. Across three generations the Pro tier therefore reads 200, 150, 273, 307 — which is not a line anybody would extrapolate.
The practical lesson is the reason this page exists in a library about inference. Generation number does not imply bandwidth, tier name does not imply bandwidth across generations, and the only reliable way to know is to read the specification page for the exact configuration. Within the M3 Max and M4 Max there are even two different figures for the same chip name, separated only by CPU and GPU core count: 300 against 400 on M3 Max, 410 against 546 on M4 Max.
The Ultra plateau
The top tier is the flattest part of the table: 800, 800, 819. Three generations of Ultra chip, and roughly two and a half per cent of bandwidth growth between the first and the last, while the memory ceiling over the same period went from 128 GB to 512 GB — a fourfold increase.
For local inference that divergence is the single most consequential fact in the table. Capacity has grown fourfold and single-stream decode speed for a model of a given size has grown by two and a half per cent, so the largest model an Ultra can hold has been getting steadily slower to decode relative to the machine’s capacity. The consequences for choosing a tier are worked through in the Mac Studio memory ceiling page.
What it means for a fixed model
Convert the table into the thing you care about. Hold the model fixed at mlx-community/Qwen3-14B-4bit, whose safetensors shards total 8,307,898,514 bytes — 8.31 GB — read through the Hugging Face model API on 11 August 2026. Decode at batch size one reads all of it per token, so the ceiling is bandwidth over 8.31 GB:
M2 100 GB/s 12.0 tok/s M3 100 GB/s 12.0 M4 120 GB/s 14.4 M5 153 GB/s 18.4 M2 Pro 200 GB/s 24.1 M3 Pro 150 GB/s 18.1 <- lower than the generation before it M4 Pro 273 GB/s 32.9 M5 Pro 307 GB/s 36.9 M4 Max 546 GB/s 65.7 M5 Max 614 GB/s 73.9 M3 Ultra 819 GB/s 98.6 upper bound only: batch one, dense, short context, and 100% of Apple's published bandwidth, which no real kernel achieves
Nothing in that table was measured. It is one division per row, and the derivation with all four of its assumptions is in unified memory bandwidth and tokens per second. For the real figure on a machine you have, run mlx_lm.generate and read the Generation: line; for a machine you do not have yet, the ratio between two rows is a better guide than either row’s absolute value, because the fraction of peak bandwidth a kernel achieves is roughly similar across the family.
The other thing the table is good for is judging a second-hand machine. An M1 Max at 400 GB/s has a higher decode ceiling for a given model than an M4 Pro at 273 GB/s, despite being three generations older; an M3 Pro at 150 GB/s has a lower one than the M2 Pro it replaced. For this specific workload the useful sort order is bandwidth first and memory tier second, and the release year carries almost no information — which is not true of anything else you would choose a Mac for.