Eight single-port 400Gb InfiniBand HDR or Ethernet CX7 adapters, one per accelerator, for scale-out; storage on a dual-port 100Gb RoCE adapter and management on gigabit Ethernet. Moore Threads publishes no scale-up fabric bandwidth for the chassis.
Moore Threads' eight-card training and inference server, and the hardware underneath the KUAE cluster product. An 8U rack chassis holds eight OAM compute modules alongside two x86 or C86 processors, up to 2 TB of host DDR5-5600 RDIMM in a typical training build of 32 x 64GB or 24 x 96GB, two 960 GB SATA boot drives and four 3.84 TB PCIe Gen4 NVMe data drives. Each accelerator gets its own 400Gb InfiniBand or Ethernet port, so scale-out runs over standard fabrics. Moore Threads rates the machine at 11.4 kW and 100 kg. The specification table identifies the accelerator only as an OAM compute module, but Moore Threads publishes the same inference benchmark footnote on two pages, describing the test hardware once as six SGX5000 eight-card servers and once as six S5000 eight-card servers, which is the vendor equating the two directly. No throughput figure is shown because Moore Threads publishes none for either the server or the S5000: its public claims are relative to unnamed international flagship parts, and the absolute TFLOPS values on the S5000 page are reference baselines for other vendors' products. Total accelerator memory is likewise absent, since the vendor states no system total and the S5000's own capacity is not published on the Moore Threads site. The 11.4 kW is a rated figure, not a measured draw.