M3 Max memory limits depend on the chip configuration

Apple has introduced M3, M3 Pro and M3 Max together, bringing a new graphics architecture to its Mac lineup. But the largest memory figure in the October 30 announcement doesn’t describe every M3 Max, let alone every M3-powered Mac.

File photo: Apple logo on an iMac, © Håkan Dahlström, 2010. CC BY 2.0, via Wikimedia Commons. Resized and cropped; this isn’t an M3 launch photograph.

The useful buying distinction is between shared graphics features and configuration-specific resources. A machine can support the same rendering techniques while offering substantially less memory capacity and bandwidth.

Check the Max configuration, not just its name

The M3 Max comes in two versions:

• 14-core CPU, 30-core GPU: 300GB/s bandwidth; 36GB or 96GB of unified memory.
• 16-core CPU, 40-core GPU: 400GB/s bandwidth; 48GB, 64GB or 128GB of unified memory.

Those combinations are corroborated by contemporary configuration reporting. The 128GB ceiling therefore requires a particular processor configuration. It isn’t simply an optional memory upgrade on the cheaper Max.

Apple’s chip announcement sets lower ceilings of 24GB for M3 and 36GB for M3 Pro. Unified memory gives the chip’s components access to a shared pool, but that doesn’t reserve the entire advertised capacity for a model or graphics application. The operating system and other software use memory too.

For someone planning a large local model or complex scene, capacity and bandwidth answer different questions: whether the working data can fit, and how quickly it can move. Neither number by itself establishes application speed, and this announcement isn’t an independent benchmark of a particular AI workload.

Dynamic Caching isn’t extra installed memory

Across the family, Apple introduces hardware-accelerated ray tracing and mesh shading, along with a feature it calls Dynamic Caching. Apple describes that feature as allocating GPU local memory in hardware according to the needs of each task, aiming to improve utilisation.

That allocation mechanism shouldn’t be confused with increasing the amount of unified memory a buyer orders. It addresses how the GPU uses resources; the configuration table still determines installed capacity and bandwidth.

Ray tracing provides hardware support for modelling light, while mesh shading changes how geometry is processed. Their presence doesn’t establish that every existing application will deliver the same benefit. Apple also lists AV1 decoding across the family, which is a playback capability rather than a promise of AV1 hardware encoding.

Read the baseline beside the performance claim

The MacBook Pro announcement gives M3 Max several different improvements against M1 Max: up to 2.5 times the rendering performance in Redshift, up to twice the speed for the cited MATLAB simulation, and up to 65% faster noise reduction in DaVinci Resolve Studio. Those are Apple’s workload-specific claims, not a universal 2.5-times multiplier.

The comparison footnote identifies a 16-inch M1 Max machine with a 10-core CPU, 32-core GPU, 64GB of RAM and an 8TB SSD. That isn’t an M2 Max baseline. Treating the headline as a one-generation improvement would change the comparison.

The same caution applies to the advertised battery maximum. Apple’s test description uses a preproduction 16-inch M3 Pro with 36GB of RAM, specified display brightness and web-browsing or 1080p movie-playback tasks. It doesn’t establish how long a 128GB M3 Max will run a sustained compute job.

Orders are open, with the MacBook Pro lineup starting to reach customers on November 7. Before choosing a configuration, match the actual CPU/GPU variant, memory option and workload to the claim being made. The family name alone leaves too much unspecified.