Happy about the simple, instructive schema of block floating point - which also makes it clear that it is not saving in every single possible case.
I am curious to see these block floating point formats being implemented by different vendors: I see an ominous "MX9" among others for AMD Versal, and on Strix Point they say they have some accelerated block floating point computations available, but the one somewhat pragmatic presentation I could find ("Leveraging the Iron AI Engine ...") only mentions int and bfloat16, and that one I assume is google's bfloat16, not block float.
Happy about the simple, instructive schema of block floating point - which also makes it clear that it is not saving in every single possible case.
I am curious to see these block floating point formats being implemented by different vendors: I see an ominous "MX9" among others for AMD Versal, and on Strix Point they say they have some accelerated block floating point computations available, but the one somewhat pragmatic presentation I could find ("Leveraging the Iron AI Engine ...") only mentions int and bfloat16, and that one I assume is google's bfloat16, not block float.