Intel Gaudi 3 AI Accelerator Technical Paper - '1677 TFlops per MME chip' prose vs 1678 in the tables

INTERNAL INCONSISTENCY IN INTEL'S OWN PAPER, recorded so nobody 'fixes' either number silently. Table 1 and Table 5 both print BF16 MME TFLOPS = 1678 for the Gaudi 3, and the records store 1678. But the MME Architecture prose says: 'housing eight Matrix Multiplication Engines (MMEs). Each of these engines is equipped with 64K Multiply-Accumulate Units (MACs), which collectively enable a peak throughput of 1677 TFlops per MME chip.' 1677 vs 1678 in the same document. The 1678 figure is the one in both tables and is the one stored; the 1677 prose figure is almost certainly a rounding-boundary artifact of 64K MACs x 2 ops x clock. Not a repo error - flagged as an upstream one.