While this paper does not directly deal with any specific product or technology, it presents an interesting idea that scientific computation can be accelerated not merely by blindly offloading computation-intensive tasks to GPUs (as is the present mindset of research community), but by intelligent use of both CPU and GPU together, since both have unique strengths.

As shown by these articles https://reference.wolfram.com/language/example/SpeedUpComputationsWithParallelGPUComputing.html and https://reference.wolfram.com/language/guide/GPUComputing.html, Wolfram language seeks acceleration using GPU. So, based on the ideas presented by this survey, even better acceleration may be possible by using computation power of both CPU and GPU and performing load-balancing or pipelining in them. This also avoids CPU idling and enables boosting even serial applications which cannot benefit from GPUs.

Still, you may kindly decide about relevance of this paper to the forum.