It might've done well as a PCIe (or thunderbolt) thing. There are definitely large scale users of scalar math that could make use of it but you'd need to build it in 5nm and dozens of cores to be even remotely competitive.