This ignores shared memory between GPU and CPU, akin to Intel iGPUs of yore. It also ignores the less efficient memory management of using 16KiB pages, unlike x86 using the more space-efficient 4KiB pages.
This is a macrumours.com article, though.
Unified memory means that you need more, not less memory. CPU, gps and neural engine uses one memory pool, so if you e.g load huge ai model, you left nothing for applications.
Sidenote: Apple pricing for extending memory is ridiculous