I wish AMD would target AI more. They could create a niche unique offering, something like a 128 GB video card with float16 support. Then optimize and open source a cuda replacement.
You address the 2TB NVMe SSG memory as a filesystem directly from the card itself - no PCIe system transfer involved. 24.6TFLOPS half-precision.
You can program using OpenCL or HIP (opensource cuda replacement )
https://pro.radeon.com/en/product/pro-series/radeon-pro-ssg/.