In short
Unified memory is a shared physical pool that multiple processors can access. For local AI, its attraction is capacity: an integrated GPU can potentially work with more model data than a small graphics card holds. Allocation and software still set limits.
What is unified memory?
The CPU and integrated accelerator share a memory system rather than relying on entirely separate physical RAM banks. Details vary by platform. The operating system, applications and model runtime still compete for capacity and bandwidth.
Unified memory vs VRAM
Dedicated VRAM belongs to a graphics card; unified memory also serves ordinary system work. Equal capacity is not equivalent performance. A large shared pool is valuable for fit, while a smaller fast card can excel on a workload that fits its memory.
How much of the memory can the GPU use on Strix Halo?
The GPU-visible amount depends on firmware, capacity and drivers. A 128GB listing does not supply 128GB exclusively for model weights. Inspect the runtime’s allocation and leave room for the operating system, context and temporary buffers.
Unified memory on Mac vs AMD
Mac workflows commonly use Metal or MLX; AMD systems use supported Vulkan, ROCm or other backends. Choose the software environment first. A similar memory number does not make applications and model formats interchangeable.
Frequently asked questions
Is unified memory the same as VRAM?
No. It can serve graphics workloads but remains a shared system pool rather than separate card memory.
What does "96GB VRAM" mean on a mini PC?
It usually describes an allocation from shared memory. Confirm the platform and settings rather than assuming separate physical graphics memory.