Likely to be bound by SGAM/GAM page contention in tempdb, but could be CPU bound. Should be easy to benchmark on your CPUs, with your tempdb files (mine differ). Duration can/should be measured in microseconds, but with millisecond accuracy. Best to iterate at least a thousand. Also test up to a hundred concurrent sessions.

But the above benchmark worries me less than a proposed use of a table variable, when the topic is performance. I am hoping you are well-versed in performance, execution plans, cardinality estimates, and table variables ;-).