Simulation Lab
Simulate Before You Scale.
Explore workload behaviour across GPU configurations, architectures and operating conditions before committing capital. This functional preview uses a transparent client-side demonstration model, not the production MeshAI simulation engine.
SimulatedIllustrativeLocal engine v0.1 · not a production benchmark
| Metric (illustrative) | A · Baseline | B · MeshAI-optimised |
|---|---|---|
| Quantisation / location | FP16 · on-premises | INT8 · on-premises |
| Model memory | 1638 GB (89% / accel) | 819 GB (85% / accel) |
| Accelerators required / available | 172 / 16 | 25 / 16 |
| Capacity utilisation | 100% | 100% |
| Accelerator utilisation | 100% | 100% |
| Throughput served / demand | 373 / 3,000 tok/s | 2,654 / 3,000 tok/s |
| Time to first token | 33,626 ms | 2,364 ms |
| End-to-end latency | 117,686 ms | 8,267 ms |
| Energy / day | 376.3 kWh (grid 346.2 · solar 0 · gen 0) | 376.3 kWh (grid 346.2 · solar 0 · gen 0) |
| Operating cost / month | $18,942 | $18,942 |
| Success rate | 12.43% | 88.46% |
| Cost per successful task | $0.00490 | $0.00069 |
| Resilience score | 34 / 100 | 92 / 100 |
| Sovereignty policy | Compliant | Compliant |
Loading visualisations…
Baseline · recommendations
- — Add 156 × NVIDIA H100 80GB to meet demand at 75% target utilisation.
- — Latency target missed — reduce output length, quantise, or place closer to users.
Configuration as enteredStatic batching assumptionNo power or connectivity awareness
MeshAI-optimised · recommendations
- — Add 9 × NVIDIA H100 80GB to meet demand at 75% target utilisation.
- — Latency target missed — reduce output length, quantise, or place closer to users.
Recommends INT8 — quality preservation must be validated per model and languageContinuous batching assumptionPower-aware scheduling of deferrable workLocal fallback during connectivity outages
Assumptions, formulas & provenance
Note · Results come from a deterministic first-order model running in your browser. They are simulated, not measured, and carry wide uncertainty (memory-bound decode approximation, fixed batching efficiencies, simplified power and outage behaviour). Quantisation recommendations must be validated for quality per model and language.
- Model memory
- params × bytes/param × 1.2 overhead + KV cache (users × tokens × 0.0001 GB × params × bytes/2)
- Per-sequence decode rate
- accelerator bandwidth ÷ weight bytes (memory-bound decode)
- Node throughput
- decode rate × min(users, batch cap) × batching efficiency
- Required accelerators
- max(memory shards, demand tok/s ÷ (per-accel throughput × 75% target))
- Latency
- network RTT + input tokens ÷ (decode × 8) + output tokens ÷ decode rate
- Energy
- accelerators × watts × (30% idle + 70% × utilisation) × PUE × 24h
- Cost / successful task
- (compute + energy cost) ÷ (requests × success rate)
- Success rate
- availability after grid, battery and connectivity outages
Constant assumptions
{
"bytesPerParam": {
"fp16": 2,
"int8": 1,
"int4": 0.5
},
"weightOverhead": 1.2,
"kvGBPerTokenPerBParamFp16": 0.0001,
"prefillSpeedup": 8,
"targetUtilisation": 0.75,
"idlePowerFraction": 0.3,
"pue": 1.4,
"dieselUSDPerKWh": 0.45,
"outageHours": 4,
"offshoreLatencyMs": 140,
"inCountryLatencyMs": 25,
"baselineBatching": 0.35,
"meshBatching": 0.65,
"solarSunHours": 5.5
}Next step
Request a Detailed Infrastructure AssessmentNeed calibrated assumptions, measured infrastructure data or a broader capacity assessment?