What would make you consider using a new sorting algo?

I've been working on a parallel integer sorting algorithm and benchmarking it against ipso, parlay's integer sort (radix ) etc.

For people who work on databases, analytics systems, HPC, compilers, or other performance-sensitive software, what feature would you need before seriously considering a new sorting implementation?

Examples:

Better throughput?
Better scaling? (on both key size and input size )
Lower memory overhead?
NUMA results?
ARM benchmarks?
Stable sorting?
Key-value sorting?
Something else?

My algo currently aims to fix the scaling issue with radix , it gets slower if you increase the key size. And providing support for 128 bits , as most parallel radix implementation only support 64 bits in a single pass (for 128 bits or higher you need to multiple stable passes.

I'd especially like to hear from people who have actually deployed or evaluated sorting implementations in production systems.

Basically the end goal is to build something that is actually useful to someone.

https://redd.it/1u3p7nd
@r_cpp