how to turn 128GB of RAM into 96GB of AI memory: the Strix Halo VRAM trick nobody explains

AMD's Ryzen AI Max+ 395 lets you hand up to 96GB of its unified memory to the GPU — enough to run a 70-billion-parameter model on a mini-PC. Here's… Aliteq Ravi Malhotra · Hardware Editor

Up to 96GB of a 128GB Strix Halo can be allocated as VRAM via AMD Variable Graphics Memory — leaving ~32GB for the OS.

It's a unified memory architecture: CPU, RDNA 3.5 iGPU and RAM share one physical pool, so 'VRAM' is just RAM you've handed to the GPU.

The split is set in the BIOS at boot, not changed at runtime — that's the step people miss.

Capacity is the win: 96GB fits Llama 70B (Q6, ~58GB), Qwen 72B (Q5, ~50GB), even Mixtral 8x22B (Q4, ~80GB) — impossible on a 24GB card.

Bandwidth is the catch: ~256 GB/s vs 1,000+ GB/s on a discrete GPU, so big models run but slower, especially on long prompts.

Strix Halo's up-to-96GB VRAM allocation is a genuine breakthrough for capacity: it runs 70B-class models — even Mixtral 8x22B — in a silent, efficient mini-PC for well under a discrete multi-GPU…

Capacity champion, bandwidth compromise

Read the full story

Aliteq how to turn 128GB of RAM into 96GB of AI memory: the Strix Halo VRAM trick nobody explains Read the full story on Aliteq