That article seems confused and/or misinformed on many points, on many levels.
As for the GP cited use case, there's no plausible way in which the compilers involved in the V8 build process would start using less memory due to their memory management machinery getting magically transformed by the hardware from tracing GC to refcounting GC, even assuming LLVM had been using tracing GC in the first place.
(Also the tracing GC memory overhead claims originating in Apple's marketing copy are way hyperbolic, modern tracing GCs don't require double storage used in "mark and sweep" GCs)
