Data table
Apple Silicon unified memory & bandwidth: M1 through M5
As of July 2026, Apple Silicon memory bandwidth spans 68 GB/s (M1) to 819 GB/s (M3 Ultra), and unified memory reaches 512 GB — but macOS only exposes roughly 67–75% of it to the GPU, which is the number that decides whether a model fits.
M-series memory specification table
Bandwidth figures are Apple's published theoretical peaks in decimal GB/s. The last column applies the default macOS Metal working-set limit (≈66.7% of total memory at 36 GB or below, ≈75% above) to the largest memory option — this, not the headline RAM figure, is what a model must fit into. Machine-readable copy: apple_silicon.json.
| Chip | Year | Peak bandwidth (GB/s) | Memory options (GB) | GPU-usable at max config (GB, default limit) |
|---|---|---|---|---|
| M1 | 2020 | 68.25 | 8 / 16 | ≈10.7 |
| M1 Pro | 2021 | 200 | 16 / 32 | ≈21.3 |
| M1 Max | 2021 | 400 | 32 / 64 | ≈48 |
| M1 Ultra | 2022 | 800 | 64 / 128 | ≈96 |
| M2 | 2022 | 100 | 8 / 16 / 24 | ≈16 |
| M2 Pro | 2023 | 200 | 16 / 32 | ≈21.3 |
| M2 Max | 2023 | 400 | 32 / 64 / 96 | ≈72 |
| M2 Ultra | 2023 | 800 | 64 / 128 / 192 | ≈144 |
| M3 | 2023 | 100 | 8 / 16 / 24 | ≈16 |
| M3 Pro | 2023 | 150 | 18 / 36 | ≈24 |
| M3 Max (30-core GPU) | 2023 | 300 | 36 / 96 | ≈72 |
| M3 Max (40-core GPU) | 2023 | 400 | 48 / 64 / 128 | ≈96 |
| M3 Ultra | 2025 | 819 | 96 / 256 / 512 | ≈384 |
| M4 | 2024 | 120 | 16 / 24 / 32 | ≈21.3 |
| M4 Pro | 2024 | 273 | 24 / 48 / 64 | ≈48 |
| M4 Max (32-core GPU) | 2024 | 410 | 36 | ≈24 |
| M4 Max (40-core GPU) | 2024 | 546 | 48 / 64 / 128 | ≈96 |
| M5 | 2025 | 153.6 | 16 / 24 / 32 | ≈21.3 |
| M5 Pro | 2026 | 307 | 24 / 48 / 64 | ≈48 |
| M5 Max (32-core GPU) | 2026 | 460 | 36 / 48 | ≈36 |
| M5 Max (40-core GPU) | 2026 | 614 | 64 / 128 | ≈96 |
Sources & date: Apple newsroom announcements and tech-spec pages — M4 Pro/Max, M3 Ultra (819 GB/s), M5 Pro/Max, MacBook Pro 16-inch (M5) tech specs. Retrieved 2026-07-13.
Unified memory changes the fit math
On a PC, model capacity is the VRAM printed on the box. On a Mac, CPU and GPU share one memory pool, and macOS reserves a slice for the system: by default, Metal's GPU working set is capped near two-thirds of total memory on machines with up to 36 GB, and near 75% above that. This is why a "16 GB" Mac behaves like a ~10.7 GB GPU for model loading — enough for an 8B model at Q4 (about 7.2 GB total, per our model VRAM table) but not for a 14B. Power users raise the cap with sudo sysctl iogpu.wired_limit_mb=…, trading system stability headroom for a few extra gigabytes.
The capacity ceiling, however, is where Macs pull far ahead. A 128 GB M4 Max exposes roughly 96 GB to the GPU — double an RTX 5090 plus an RTX 4090 combined — and the 512 GB M3 Ultra can hold gpt-oss-120b (≈65 GB), Llama 3.3 70B at Q8 (≈75 GB), or several large models simultaneously. For models in the 40–120 GB range, a high-memory Mac is often the only single-box way to run them at all without a multi-GPU rig.
Bandwidth tiers, not generations, set the speed
Within Apple's lineup, generation matters far less than tier. Every base chip (M1 → M5) sits between 68 and 154 GB/s; every Pro chip between 150 and 307 GB/s; Max chips between 300 and 614 GB/s; Ultras at 800–819 GB/s. Because token generation streams the whole active model per token, measured llama.cpp throughput tracks these tiers almost linearly — an M2 Pro (200 GB/s) generates about 38 t/s on a 7B Q4 model while a same-year M2 Max (400 GB/s) reaches about 63 t/s. The measured tokens/sec table lists the full Apple set alongside discrete GPUs.
Two quirks are worth flagging when reading the table. The M3 Pro was a bandwidth regression (150 GB/s vs. 200 GB/s for the M1/M2 Pro), making some 2023 Macs slower for LLM work than their 2021 predecessors. And within a single chip name, GPU-core count can change the memory system: the 40-core M4 Max carries 546 GB/s versus 410 GB/s for the 32-core variant — a 33% generation-speed difference hidden behind one "M4 Max" label. Apple applies the same split to the M5 Max (614 vs. 460 GB/s), so check the exact configuration, not just the chip family.
Frequently asked questions
How much unified memory can the GPU actually use on a Mac?
By default macOS caps the GPU working set at roughly two-thirds of total memory on Macs with 36 GB or less, and about 75% above that. A 16 GB Mac therefore exposes about 10.7 GB to Metal, and a 64 GB Mac about 48 GB. The limit can be raised with the iogpu.wired_limit_mb sysctl at your own risk.
Which Apple chip has the most memory bandwidth?
The M3 Ultra, at 819 GB/s, followed by the M1 Ultra and M2 Ultra at 800 GB/s. Among current laptop chips the M5 Max (40-core GPU) leads at 614 GB/s. For comparison, an RTX 4090 has 1,008 GB/s.
Can a MacBook run a 70B model?
Yes, with 64 GB or more of unified memory. Llama 3.3 70B at Q4_K_M needs about 46 GB in total at an 8k context, which fits within the ~48 GB GPU share of a 64 GB M4 Max or M5 Max. Expect roughly 7 to 9 tokens/sec on a 400-546 GB/s chip.
Is unified memory as fast as GPU VRAM?
Usually not. Base M-series chips (100-154 GB/s) are far below any discrete GPU, Pro chips (200-307 GB/s) are comparable to entry discrete cards, and only Max/Ultra chips (400-819 GB/s) reach discrete-GPU territory. The advantage is capacity: up to 512 GB in one pool, which no consumer GPU approaches.
Why is a base M3 slower than an M1 Pro for local LLMs?
Bandwidth. The M3 has 100 GB/s versus the M1 Pro's 200 GB/s, and token generation speed scales with memory bandwidth rather than chip generation. Measured llama.cpp results show the 2021 M1 Pro generating roughly 45% faster than the 2023 base M3 on the same 7B model.