I have SQL Server 2019 CTP 3.2 installed on my virtual machine, which has 8GB of memory (max server memory set to 6 GB) and 4 vCPUs. I created four (4) tempdb data files, each sized to 1GB.
I restored a copy of WideWorldImporters and then created three stored procedures (definitions below). Each stored procedure accepts a date input, and pushes all rows from Sales.Order and Sales.OrderLines for that date into the temporary object. In Sales.usp_OrderInfoTV the object is a table variable, in Sales.usp_OrderInfoTT the object is a temporary table defined via SELECT … INTO with a nonclustered added afterwards, and in Sales.usp_OrderInfoTTALT the object is a pre-defined temporary table which is then altered to have an additional column. After the data is added to the temporary object, there is a SELECT statement against the object that joins to the Sales.Customers table.
This isn’t a be-all, end-all performance test, but it does give us an idea of what can work better with memory-optimized tempdb metadata.