It's an interesting post and an interesting area of discussion. To some extent this is how people see GPUs: you balance your size and numbers of CPUs and GPUs in your system to match your workload. We did a lot of work at Codeplay to enable people to write code once and it adapts to different hardware, but the hardware designers were resistant. This is true for other software developers working in the field. There are various reasons for this, but it restricts the way you can realistically map software to hardware. So instead, what actually happens in most of the semiconductor industry is what has previously been highly successful in signal processing: hardware designers optimize a hardware/software system for a specific set of workloads. What you're doing is gaining efficiency at a cost of longer time to market. That approach is highly successful in communications systems. But it's been a poor tradeoff in AI, where very short time to market is the top priority. It would be great to fix it, but it's mostly blocked by industry resistance: the software techniques to do it are well known and understood by many HPC software experts. These techniques are very unpopular in hardware design though. So it's a people problem