In addition to being an “in-memory” technology, Analysis Services Tabular is also a “column-store” technology which means all the values in a table for a single column are stored together. As a result – and this is especially true for dimensional models – we are able to achieve very high compression ratios. On average, you can expect to see compression ratios anywhere from 10x-100x depending on the model, data types, and value distribution.
What this ultimately means is that your 2 TB data mart will likely only requrie between 20 GB of memory (low-end) and 200 GB (high-end) of memory. That’s pretty amazing – but still leaves us with a fairly wide margin of uncertainty. In order to further reduce the level of uncertainty, you will want to take a representative sample from your source database, load it into a model on a server in your DEV environment, and calculate the compression factor.
Read the whole thing; Bill has several factors he considers when sizing a machine.