You can probably just dedicate 1GB for the framebuffer, and then let Linux dynamically allocate memory to it at runtime. As far as I can tell this doesn't impact performance, so there's no downside. (Older AMD stacks required a static partitioning under Linux, I think).
I haven’t had much time with it, but I’ve had to set the split in the BIOS. There’s probably a way to do it from within Linux though. Also hoping some progress is made on using the AMD NPU in Linux. I know it only recently got kernel level support.
I thought I had to split it in the BIOS, but then I just didn't (this is on a 2025 machine), and llama ended up with the same available "GPU" ram either way (confirmed by running inference on it).
You can probably just dedicate 1GB for the framebuffer, and then let Linux dynamically allocate memory to it at runtime. As far as I can tell this doesn't impact performance, so there's no downside. (Older AMD stacks required a static partitioning under Linux, I think).
I haven’t had much time with it, but I’ve had to set the split in the BIOS. There’s probably a way to do it from within Linux though. Also hoping some progress is made on using the AMD NPU in Linux. I know it only recently got kernel level support.
I thought I had to split it in the BIOS, but then I just didn't (this is on a 2025 machine), and llama ended up with the same available "GPU" ram either way (confirmed by running inference on it).
Oh that's fantastic, I'll give it a try. thank you!