# Local LLM hardware sizing worksheet

Editable planning and measurement record. No benchmark, hardware minimum, price or product winner is supplied.

Workflow/quality reviewer: [value]. Expected useful output: [definition]. Runtime/build: [value]. Model/revision/hash/licence: [values]. Quantisation: [value]. Context/input/output allowance: [tokens]. Tokenizer: [version]. Active overlapping requests and arrival pattern: [values].

## Memory and environment

Candidate hardware, OS, drivers and runtime support: [values]. Workstation or dedicated server: [value].

| Item | Host RAM estimate/measured | Discrete accelerator memory estimate/measured | Shared pool, if applicable | Evidence/assumption |
| --- | --- | --- | --- | --- |
| Resident model weights | To record | To record | Count once | To record |
| Active context/cache | To record | To record | Count once | To record |
| Runtime buffers | To record | To record | Count once | To record |
| Other resident services/desktop work | To record | To record | Count once | To record |
| Agreed reserve | To agree | To agree | Count once | Rationale |

Do not add a shared-memory pool to itself. Record actual CPU/GPU placement and offload. Include storage, cooling, power and recovery requirements separately.

## Acceptance workload

Fixtures/rubric/version: [reference]. Repetitions: [value]. Accepted quality: [rule]. Time-to-first-useful-output and completion thresholds: [values]. Overload/recovery rule: [value]. Normal desktop activity included: [value].

| Scenario | Requests completed/attempted | Quality/review result | First useful output/completion distributions | Peak memory/placement | Failures/recovery |
| --- | --- | --- | --- | --- | --- |
| Cold start | Not measured | Not measured | Not measured | Not measured | Not run |
| Warm single user | Not measured | Not measured | Not measured | Not measured | Not run |
| Intended overlap | Not measured | Not measured | Not measured | Not measured | Not run |
| Overload burst then normal load | Not measured | Not measured | Not measured | Not measured | Not run |

Include queue waiting, failures and cancellations. Whole-machine electricity cost needs an appropriate host-level measurement; accelerator reporting alone is insufficient.

## Decision and cost

Reuse/buy/reduce scope/retain hosted or manual route: [decision and evidence]. Quote date/currency/tax treatment: [values]. Compare purchase/allocation, energy, integration, storage/backups, maintenance, recovery and review over the same period. Mark unknowns explicitly.

Configuration owner and known-good rollback: [values]. Local data-path/network test: [separate evidence, not established by hardware benchmark]. Supplier demonstration must use this workload, not a substitute generic benchmark.