Throughput ? Estimated TeraFLOPs/second. Based on TPU count, float type (lower precision = higher potential FLOPs), interconnect speed, torus knot twist (minor impact), and scaling efficiency (drops slightly with more TPUs). Includes minor random variation. : 0 TFLOPS

Efficiency ? Percentage of theoretical peak performance achieved. Primarily affected by interconnect speed and scaling efficiency (communication overhead increases with size). Torus knot twist has a minor negative impact. Base is 95%, capped between 50-98%. Includes minor random variation. : 0 %

Est. Latency ? Estimated communication latency in milliseconds. Increases logarithmically with TPU count (more hops). Decreases with better interconnects. Torus knot twist slightly reduces latency (potentially shorter paths). Base is 5ms. Includes minor random variation. : 0 ms

Est. Cost Factor ? Abstract relative cost. Primarily driven by the number of TPUs. Better interconnects add a moderate cost increase. Lower precision float types (FP8/INT8) slightly reduce cost factor, FP32 slightly increases it. Includes minor random variation. : 0