It’s quite common for there to be a latching bottleneck in tempdb that can be traced back to temporary table usage. If there are lots of concurrent connections running code that creates and drops temporary tables, access to the database’s allocation bitmaps in memory can become a significant bottleneck.
This is because only one thread at a time can be changing an allocation bitmap to mark pages (from the temp table) as allocated or deallocated, and so all the other threads have to wait, decreasing the workload throughput. Even though there has been a temporary table cache since SQL Server 2005, it’s not very large, and there are restrictions on when the temporary table can be cached (e.g. only when it’s less than 8MB in size).
This is great advice; read the whole post.