Hardware
CPU, GPU, VRAM, RAM, storage, OS, power, cooling, and any unusual constraints.
Home / Benchmarks
Lab logTransparent format
Local AI benchmark queue and test methodology.
A public plan for TokenByte hardware and workflow tests: exact machine, model, settings, result, limitation, evidence, and buying verdict.
Current evidence status: the rows below are queued tests, not completed benchmark results. TokenByte should not use them as performance claims until the matching evidence files, screenshots, settings, and measured values are published.
Public test queue
What is planned, what is measured, and what is still missing.
This page separates planned tests from completed evidence. A queued row is a promise of method, not a performance claim.
Structured benchmark queue
The public data file keeps the queue honest: each test has a status, required evidence, target metrics, and the guide it will update.
| Test | Machine | Workload | Metrics to publish | Status | Evidence needed |
|---|---|---|---|---|---|
| RTX 3090 ComfyUI workflow | RTX 3090 / 24GB VRAM | Image generation, upscale chain, batch output | VRAM use, render time, resolution, workflow file, failure notes | Queued | Workflow screenshot, exact graph, seed/settings, VRAM capture, output sample, failure notes |
| Mac Mini local model run | Mac Mini + external SSD | Summaries, transcripts, small local LLM prompts | Tokens/sec, memory pressure, model name, prompt set, output quality | Queued | Model name, quantization, app version, prompt set, memory capture, output notes |
| Storage stress test | NVMe / external SSD | Model library loading, output folder growth, backup flow | Capacity used, load time, heat, reliability notes | Queued | Drive model, enclosure, folder size, transfer notes, heat notes, backup result |
| Automation folder watcher | Mac Mini or local PC | File summaries and Markdown output | Files processed, runtime, error rate, model cost/privacy notes | Queued | Script/config, file count, runtime log, error cases, before/after output sample |
Benchmark fields
Every result needs receipts.
TokenByte benchmark posts should include enough context for a reader to reproduce or reject the result.
Gear hub
Software
Model names, app versions, drivers, workflow files, quantization, and settings.
Run Data
Speed, memory, VRAM, temperature, output size, time-to-result, and failure modes.
Verdict
Who should buy, who should skip, cheaper alternatives, and what to upgrade first.
Measurement protocol
How a queued row becomes a published result.
The page is allowed to influence buying advice only after the test moves from queued to measured in the public JSON log.
Minimum evidence gate
- Record the exact hardware: CPU, GPU, VRAM, RAM, storage, OS, driver, cooling, and power constraints.
- Record the exact software: app version, model or workflow name, quantization, settings, seed when relevant, and prompts or input files.
- Run the same workload at least three times when the result is a speed, time, or memory claim.
- Publish the failure notes, not only the best run. A setup that crashes, swaps, overheats, or silently degrades output is not a clean win.
- Attach screenshots, output samples, workflow files, logs, or photos before changing a buying verdict.
| Status | Allowed claim | Not allowed yet |
|---|---|---|
| Queued | TokenByte plans to test this workload and has listed the required evidence. | Speed claims, winner language, product rankings, or completed-review scores. |
| Measured | The row has original measurements and evidence files attached to the public log. | Broad claims outside the tested hardware, model, settings, or workflow. |
| Retest | Older data exists, but the recommendation needs a new run before changing buying advice. | Using stale numbers as current buying proof. |