Decentralized AI Compute Dispatch & Network Latency Simulator

Model & Hardware Parameters

Model Parameters (B) 70B

Micro-Batch Size 8

Pipeline Shards (Layers Sharded) 3 Nodes

Selected Cluster Specs

US-East (Primary) us-east

Cluster Compute (TFLOPS) 3200

Inter-Region Bandwidth (Gbps) 100

Network Latency (ms) 10

Pipeline Step Time

142 ms

Bubble: 24.5%

Cluster Throughput

4,820

TFLOPS Aggregate

Comm / Compute Ratio

0.82

Bound: Network Latency

Micro-batch Pipeline Schedule Execution Plan: Step 0/8 Idle

Pipeline Analysis

Primary Link Bottleneck

EU-West → Asia-East

Saturation: 88% @ 160ms Latency

Tensor Transfer Overhead

64.2 ms

Per Micro-Batch Transfer

Effective TFLOP Efficiency

75.5%

Hardware Compute Utilization