AMD Debuts Ryzen AI Max Series "Strix Halo" SoC: up to 16 "Zen 5" cores, Massive iGPU

[TPUnique]

Please test the HP Z2 Mini G1a, could be a fantastic machine for ML workloads @ max RAM.

 

[trsttte]

Since RDNA 3.5 is mobile, and most of the stuff comes with NPUs, does RDNA 3.5 even have AI Accelerators still? If so, what is the point?

It depends on what you consider AI accelerators on a GPU. AMD doesn't have something like tensor cores, they mention on some materials having some AI accelerators per compute unit but I don't know what that means, probably just dp4a implementation and the like. On the other hand NPUs use a different instruction set (which one? beats me) so they can't simply replace the GPU implementing the basic stuff like dp4a.

Maybe someone more knowledgeable can shed some light on this, in my opinion the best case scenario would for this RDNA3.5 for the better or worse to be just like a regular GPU and have the NPU as an extra to meet the "copilot pc" bs requirements (it's cool to have but not because of microsoft copilot pc requirements)

 

[igormp]

It depends on what you consider AI accelerators on a GPU. AMD doesn't have something like tensor cores, they mention on some materials having some AI accelerators per compute unit but I don't know what that means, probably just dp4a implementation and the like. On the other hand NPUs use a different instruction set (which one? beats me) so they can't simply replace the GPU implementing the basic stuff like dp4a.

Maybe someone more knowledgeable can shed some light on this, in my opinion the best case scenario would for this RDNA3.5 for the better or worse to be just like a regular GPU and have the NPU as an extra to meet the "copilot pc" bs requirements (it's cool to have but not because of microsoft copilot pc requirements)

NPUs are meant to do basic inference with quantized models, often INT4 or INT8, and don't have much extra capabilities apart from that.
The extra "AI accelerators" in GPUs (like XMX in Intel, Tensor cores in Nvidia, or the WMMA instructions in RDNA) are meant to do large matmuls similar to the ones in NPUs, but faster and with many different data types, such as FP16, FP8, BF16, FP32, etc etc, which allows for higher performance, quality and also for training models.