If your MacBook has 36GB of memory or less, then the GPU can utilize 66% or 2/3rds of your total memory. If, however, your MacBook has 36GB or more memory, then it can utilize 75% or 3/4ths of your total memory.
This is just a software limit, so when I want to run AI models on my MacBook, I just increase the allocation using Terminal.
Dedicated GPU memory is fixed. It can't be altered at all and cannot be used for other tasks either. For example, in MacBooks, if the GPU is idle, the CPU can utilize almost 100% of the total memory because they share a common memory pool, but in a dedicated GPU setup, if the GPU is idle, then the fixed VRAM memory (8GB, 12GB, 16GB, etc.) cannot be utilized by any other component of the PC.
Also, in a dedicated GPU setup, the system RAM and VRAM are physically separated. If a CPU loads a large dataset, it has to copy it across the motherboard to the GPU. Because MacBooks have a shared memory pool for both the CPU and the GPU, no copying is required, which makes dataset handoffs much faster than in a dedicated GPU setup.
If your MacBook has 36GB of memory or less, then the GPU can utilize 66% or 2/3rds of your total memory. If, however, your MacBook has 36GB or more memory, then it can utilize 75% or 3/4ths of your total memory. This is just a software limit, so when I want to run AI models on my MacBook, I just increase the allocation using Terminal.
Dedicated GPU memory is fixed. It can't be altered at all and cannot be used for other tasks either. For example, in MacBooks, if the GPU is idle, the CPU can utilize almost 100% of the total memory because they share a common memory pool, but in a dedicated GPU setup, if the GPU is idle, then the fixed VRAM memory (8GB, 12GB, 16GB, etc.) cannot be utilized by any other component of the PC. Also, in a dedicated GPU setup, the system RAM and VRAM are physically separated. If a CPU loads a large dataset, it has to copy it across the motherboard to the GPU. Because MacBooks have a shared memory pool for both the CPU and the GPU, no copying is required, which makes dataset handoffs much faster than in a dedicated GPU setup.
You can override the standard allocation with:
sudo sysctl iogpu.wired_limit_mb=…