The point of this Document is to define some high level targets and processes for creating the Reference Implementation.
These are a guess at what specs would be good for the device. Assumptions and guesses will be validated and corrected over time.
BFloat16 Mult, Float32 Add
Expected: 32 Virtual Units
Range: 16 to 64 Virtual Units
Expected: 100 Units per Early Exit
Range: 10 to 300 Units per Exit
Expected: 16,384 Units per Chip
Range: 8,192 to 65,536 Units per Chip
Expected: Unknown (but use 1Ghz to simplify calculations)
Range: 700Mhz to 1.3 Ghz (based on known TPU values)