Intel Gaudi 3 technical paper Table 1 - 'Networking (bidirectional) 600 GB/s (Gaudi 2) / 1200 GB/s (Gaudi 3)', which is the only directional statement Intel makes for the aggregate and which reconciles exactly with the 24x 200 GbE port count

VERBATIM from https://cdrdv2.intel.com/v1/dl/getContent/817486 Table 1, 'Intel Gaudi 2 to Intel Gaudi 3 AI Accelerator Feature Comparison': 'Networking (bidirectional) 600 GB/s 1200 GB/s' The same row set gives, for contrast, the host interface in the OPPOSITE convention: 'Host Interface Peak BW 64 GB/s (32 GB/s per direction) 128 GB/s (64 GB/s per direction)'. Intel therefore labels the networking row 'bidirectional' and the PCIe row per-direction in one table, which settles both conventions from the vendor rather than by inference. THE RECONCILIATION IS EXACT AND IT IS WHAT MAKES THIS ROW USABLE. 24 ports x 200 GbE = 4,800 Gb/s = 600 GB/s if the per-port rate is read per direction, so 1200 GB/s 'bidirectional' is precisely 24 x 200 GbE counted in BOTH directions. Three independent parts of the paper agree: the 24x 200 GbE port count (Table 4), the 48x 112G PAM4 SerDes behind it (Table 2), and this 1200 GB/s row. A reader who halves 1200 to 600 and then also halves the 200 GbE port rate is double-counting the same 2x. THE GAUDI 2 NUMBER IS THE USEFUL COROLLARY: 600 GB/s bidirectional is exactly 12 x 200 GbE per direction, so Gaudi 2 carries 24 NIC ports at 200 GbE where Gaudi 3 carries the same 200 GbE port at twice the count. The generation-on-generation 2x networking gain is a PORT-COUNT gain at a constant per-port rate, not a faster link - which matters if you are counting SerDes or ports on a board. Note this row is a whole-card NETWORKING aggregate and does NOT itself distinguish scale-up from scale-out; that split is Table 4 (w7f-intel-gaudi3-paper-t4-scaleup-vs-scaleout-port-split). The two are consistent: 21 scale-up + 3 scale-out = 24.