Huawei Atlas 650E page, Product Features prose bullet 'Provide 13.4 TB/s UB Link bandwidth per server (16-NPU FM)' - the vendor's own UNIDIRECTIONAL restatement of its own bidirectional table figure, which settles the UB Link direction convention without inference

FETCHED 2026-10-05 from https://www.hiascend.com/en/hardware/ai-server, 3,821 characters retrieved successfully. THE BULLET, verbatim from the Product Features section headed 'Ultra-high UB Link Bandwidth / 16NPUs· Ultra-high interconnect bandwidth': 'Provide 13.4 TB/s UB Link bandwidth per server (16-NPU FM)' 'Support 16-NPU UB full-mesh interconnect across 2 servers' THE SPEC TABLE on the same page gives, under 'UB Link Bandwidth': '8-NPU Server: 8 x 784 GB/s (bidirectional)' '16-NPU Full-mesh: 2 x 8 x 1.68 TB/s (bidirectional)' and under 'Networking': 'UBoE: Providing 400 Gbps per NPU (unidirectional)' 'RoCE: Providing 400 Gbps per NPU (unidirectional)' WHY THIS SOURCE IS WORTH ITS OWN RECORD. 2 x 8 x 1.68 = 26.88 TB/s, and 26.88 / 2 = 13.44 TB/s, which is the 13.4 TB/s Huawei prints in prose. HUAWEI THEREFORE RESTATES ITS OWN BIDIRECTIONAL TABLE FIGURE AS A UNIDIRECTIONAL PROSE FIGURE ON THE SAME PAGE. That is a vendor-internal confirmation of the direction convention and it is a much stronger basis for halving the table figure than any inference a reader could make - it is the vendor doing the halving itself, and doing it in the same document. It also removes the main residual risk on interconnect/huawei-ub-link: previously the 392 GB/s per direction figure depended on trusting the word 'bidirectional', and a record that had to hedge about whether Huawei meant what it said. Now the hedge is unnecessary. THE SAME PAGE ALSO DEMONSTRATES THAT THE VENDOR USES BOTH CONVENTIONS DELIBERATELY: 'bidirectional' for UB Link and 'unidirectional' for UBoE and RoCE, on one page, in adjacent rows. A vendor applying a single sloppy convention would use one word throughout; using two, correctly, on adjacent rows is strong evidence that neither word is an accident. The practical consequence is that the scale-out figures (400 Gbps per NPU = 50 GB/s) must NOT be halved a second time, while the UB Link figures must. Also on this page, for the record: 'NPU | 8 × 950DT NPU', 'On-chip Memory | 8 × 96 GB total, 4.0 TB/s peak memory bandwidth', 'Bus Type | PCIe 5.0 x 5', 'Power Usage | ～14.5 kW', and the compute rows 12.48 PFLOPS mxFP4 / 6.43 PFLOPS mxFP8-FP8-HiF8 / 3.40 PFLOPS FP16-BF16. The sibling SuperPoD page (cn-huawei-atlas-950-superpod-spec) gives '64 × 1.68 TB/s (bidirectional)' for the 64-NPU cabinet and '1.72 PB/s UB Link bandwidth' at 1024 NPUs, plus 'Support 256 TB unified global memory addressing & memory semantics' and 'Cross-cabinet all-optical interconnect with 2+2 optical path protection'.