Optional · Step 0 · Quick check
Optional first pass with free calculators — KV cache, GPU memory, cost, throughput, carbon. Under 30 seconds each before you commit to a full analysis.
These are optional. The platform workflow starts when you define a workload in Lens, benchmark in Pulse, and export a decision report. Calculators are for 30-second ballpark checks.
Estimate KV cache memory from model, context, batch, and precision.
Break down weights, KV cache, activations, and total HBM required.
Estimate GPUs needed for a target QPS, prompt, and output length.
Daily, monthly, and annual GPU spend by type and utilization.
Theoretical decode throughput from roofline model.
Memory delta when scaling context from one length to another.
Active vs total parameters, FLOPs, and routing overhead.
Suggest TP/PP/EP split for a model on available GPUs.
Estimated collective communication time per token.
Whether MoE collectives dominate the critical path.
Time to first token from prefill roofline.
How batch size trades latency for throughput.
Compare memory and throughput across precisions.
power, power_streaming, energy, co2, carbon, and temperature from GPU utilization.
Grade your inference deployment on carbon intensity per 1M output tokens.
Infrastructure cost amortized across daily active users.
Quick scorecard for deployment configuration gaps.
Predict whether workload is compute-, memory-, or network-bound.
Recommend TP, PP, EP, and DP strategies.
Suggest serving topology for your workload shape.
Shortlist models by latency, cost, and context constraints.