Sept 17 — At the Ascend AI industry summit where the Ascend 950 super-node was unveiled, Digital China subsidiary KunTai launched a new Ascend 950-based AI compute product matrix: KunTai A989 I5 standard server, KunTai A989 I5 air-cooled super-node, and KunTai PoD2000 A5 liquid-cooled super-node. The range targets large-model training and high-concurrency inference. The KunTai PoD2000 A5 integrates 1,024 cards per super-node, supports 256 TB unified memory addressing, and reports a thousand-card

Sept 17 — At the Ascend AI industry summit where the Ascend 950 super-node was unveiled, Digital China subsidiary KunTai launched a new Ascend 950-based AI compute product matrix: KunTai A989 I5 standard server, KunTai A989 I5 air-cooled super-node, and KunTai PoD2000 A5 liquid-cooled super-node. The range targets large-model training and high-concurrency inference. The KunTai PoD2000 A5 integrates 1,024 cards per super-node, supports 256 TB unified memory addressing, and reports a thousand-card MTBF of 300 hours, claiming a performance lead of more than 25% versus peers. It uses TB-level NPU interconnect bandwidth and ~3 μs RTT to alleviate intra-cluster compute and storage communication bottlenecks.