Apple
macOS lets the GPU keep a limited share of memory wired for its use: Metal reports it as the recommended maximum working set. Apple does not publish the share. Owner reports put it at two thirds of memory on smaller Macs and about three quarters on larger ones, so on a 128 GB Mac the default is about 96 GB. The MLX documentation shows how to raise it with sudo sysctl iogpu.wired_limit_mb=N. Pages here show both the default and a raised limit of total memory minus 8 GB for macOS, each marked as an assumption.
AMD Ryzen AI Max
AMD’s Variable Graphics Memory setting reserves up to 96 GB of a 128 GB machine for the GPU on Windows. AMD recommends its “Custom (75%)” setting for AI work, and for 32, 64 and 192 GB machines, where it gives no figure, this site applies that 75%. On Linux the GPU can borrow system memory through GTT, which defaults to about half of RAM and can be raised; owners running large models usually do.
NVIDIA GB10
The DGX Spark and its OEM versions share 128 GB coherently between the Grace CPU and the Blackwell GPU, and NVIDIA publishes no GPU allocation limit. Pages here assume total memory minus 8 GB for the operating system and label it.
Why it matters
The difference between 96 and 120 GB is the difference between a 100B-class MoE model at 4-bit with room for context and not running it at all. The fit tables use the first figure listed for each machine and show a separate “fits with raised limit” status where raising the limit is what makes a model fit.