256 deployed GPUs · 72-node-standard network baseline · 204.8 Tb/s live InfiniBand compute fabric
Primary report now separates the 32-node deployed compute scope from the 72-node-standard network baseline. Covers split BOM, dual UFM, SN5610/SN4700/SN2201 layers, and the revised 1.5 MW versus 680 kW framing.
Updated architecture summary treats N1-N6 as the refreshed network/services placement envelope rather than a simple N1/N2 occupied story. Covers live Q3400 compute fabric, retained Spectrum-4, dual UFM, and the added SN5610/SN4700/SN2201 layers.
Per-server deep-dive: ASUS XA NB3I-E12 BOM, CX8 rail mapping, BF3 wiring, X710 management mapping into the SN2201 layer, and server-side power audit for the live 32-node compute layer.
High-level summary for the current package: 32 deployed nodes, 72-node-standard network baseline, 34 contracted racks, 1.5 MW facility allocation, and split baseline-versus-deployment reporting rules.
Cooling stays anchored to 650+ kW with 4 total CRAC/CRAH units and N+1 coverage, while power framing is updated to 1.5 MW facility allocation and a 680 kW current rack ceiling. Supporting systems now assume the refreshed SN2201/SN4700/SN5610 management and service layers.