Calculator

Inference Calculator

Estimate OOM risk and memory requirements. Compare baseline engines vs S88 (expected).

Inputs

Use presets or enter custom details. Estimates are directional.

GPU (preset)

Baseline engine

GPU VRAM (GB)

System RAM (GB)

Fast SSD cache (GB)

Model (preset)

Quantization / dtype

Model size (B params)

Context length

Concurrent sessions (decode batch)

Serving setup

KV cache dtype

Prefill batch (prompt stage)

KV window (optional)

Tensor parallel (TP)

Pipeline parallel (PP)

CUDA graphs (adds VRAM overhead)

Paged allocation (reduces fragmentation) Static KV prealloc (wastes VRAM)

Prefix sharing (SGLang)

Shared prefix KV: 0%

Business impact (optional)

These are simple inputs for ROI / Opex / Capex math. Leave blank to ignore.

Nodes

Electricity ($/kWh)

S88 price ($/node/mo)

Baseline power (W)

S88 power (W)

Capex upgrade avoided ($/node)

Downtime cost ($/hour)

Baseline outage (hrs/mo)

S88 outage (hrs/mo)

No data leaves your browser.

Baseline

Engine: vLLM

—

Per‑GPU VRAM (range) —

Your GPU VRAM —

KV cache (range) —

Weights (per‑GPU) —

Parallelism —

This estimates “all hot” residency in VRAM (typical baseline behavior).

With Sector88 (expected)

Memory-aware orchestration

—

VRAM (range) —

RAM (range) —

SSD cache (range) —

VRAM relief —

Recommended next step

—

Sector88 is built around crash-avoidance and operational controls. Advanced tiering/predictive policies are roadmap items.

ROI summary (from your inputs)

This is optional math using your assumptions (energy, downtime, capex, subscription).

Annual energy savings —

Annual downtime savings —

Capex avoided (one-time) —

Total annual savings (opex) —

Total monthly savings (opex) —

Payback vs S88 price —

Want measured numbers (not estimates)?

We can enable the benchmark runner (Runtime + Hub) on your hardware and generate a report from real telemetry.

Request Benchmark Access

Uses the standard KV-cache sizing formula (layers × KV heads × head dim × sequence length × batch × dtype), plus conservative overhead ranges. Validate with a real benchmark before production deployment.

Inference Calculator

Inputs

Serving setup

Business impact (optional)

Download sizing report