I'm a little confused by the concept of granularity. I appreciate you taking the time to answer these few questions for me.
Q1: Can granularity[0] be greater than native_granularity[0] ? ( And can granularity[1] be greater than native_granularity[1] ? )
I have an example here that I'd like you to take a look at.
Q2: For this eample, if memory is sufficient and native_granularity is [128, 128], can I fuse all the nodes and use granularity [128, 128, 32] ?
Q3: If Q2 is feasible, Is the computing time for a single turn calculated as $2000 \cdot \frac{256}{128} + 2000 \cdot \frac{32}{256} = 4250$?
I'm a little confused by the concept of granularity. I appreciate you taking the time to answer these few questions for me.
Q1: Can granularity[0] be greater than native_granularity[0] ? ( And can granularity[1] be greater than native_granularity[1] ? )
I have an example here that I'd like you to take a look at.
Q2: For this eample, if memory is sufficient and native_granularity is [128, 128], can I fuse all the nodes and use granularity [128, 128, 32] ?
Q3: If Q2 is feasible, Is the computing time for a single turn calculated as$2000 \cdot \frac{256}{128} + 2000 \cdot \frac{32}{256} = 4250$ ?