Unified memory accelerators like GH200, GB200 and even GB300 have demonstrated uses and benefits in high performance scientific computing. They even appear in the top 10 of Top500.

It is also pretty much accepted in the HPC community that accelerators are necessary, but the complexity of heterogeneous programming, necessity of careful design to avoid data transfer bottlenecks and the cost of accelerators are preventing commonplace adoption.

However, even though newer CUDA versions and devices like DGX Spark are addressing these concerns to significant degree, I haven't seen any benchmarks, any results or any declaration of interest from the community. If you are part of the HPC community, please share your opinions or experience.