[Updated 4-15-2013]
I've come up with a solution for ESRI JSAPI 3.4 and AMD.
As your JavaScript projects get more and more complex, loading all of those Dojo classes can really slow down your load time. All those dojo.require calls add up in a hurry. The Dojo Build System can be a huge help in speeding up the load time and general performance of your apps. For example, a build that I ran on a recent project took the number of JavaScript requests on page load from 53 down to 5. The css request went from 16 down to 4. This ended up cutting the load time in half! Other nice features include stripping out all of the console calls, minifying your JavaScript and interning all of your widget templates.
The rest of this post assumes some familiarity with the Dojo Build System. If you haven't looked at it before, the documentation is worth reading. There's even a fancy new tutorial.
After reading all of the Dojo documentation it's easy to get excited about the possibilities. However, you will quickly find that mixing the ESRI api into the equation makes a big mess of everything. For example, the dojo build system assumes that you are hosting everything yourself. But because ESRI has not released a source/unbuilt version of their api that we can download we are stuck loading Dojo from their servers. The other problem is that when you load the ESRI api you are really loading their layer file which can have a lot of overlap with your layer file thus adding a lot of duplicate code. Not to mention the problems that the build system has when it sees: dojo.require("esri..."); and it doesn't know where to get it. Over the last few months I've developed a solution to overcome these problems and end up with a lean and mean (for the most part) product in the end.
