Verification workload

The TJ
Benchmark.

Every card that wants to run creator jobs proves itself the same way: one fixed image fine-tune, timed and measured. It is adapted from NVIDIA's own open-source example, the one that teaches Stable Diffusion XL to draw Toy Jensen. That example is where this project got its name.

A data-centre GPU board under green light, generating a grid of images

Where this comes from.

NVIDIA publishes an AI Workbench example project that fine-tunes SDXL on a handful of photos of their toy mascot. It is public, it is Apache 2.0, and it is a real, repeatable GPU workload. We took it as the yardstick.

github.com/NVIDIA/workbench-example-sdxl-customization

NVIDIA AI Workbench: SDXL Customization. Fine-tunes Stable Diffusion XL with DreamBooth so it learns one specific subject. NVIDIA's own example subject is Toy Jensen, the toy figure of their founder that they use as a demo character.

Apache 2.0Stable Diffusion XLDreamBoothNVIDIA AI Workbench
# the upstream project, unchanged git clone https://github.com/NVIDIA/workbench-example-sdxl-customization # then open in NVIDIA AI Workbench and run FineTuning-SDXL.ipynb

Facts on this page are read from the repository's own README. Verified 10 Sep 2026.

What the upstream project specifies

FieldValue
Base modelStable Diffusion XL
MethodDreamBooth fine-tune
Example subjectToy Jensen
NotebookFineTuning-SDXL.ipynb
GPUAny NVIDIA training GPU
Tested on1× NVIDIA A100-80GB
Storage40 GB
Operating systemUbuntu 22.04
LicenceApache 2.0

What we added on top.

The upstream notebook proves a model can learn a subject. We needed it to prove a machine is real, present and as fast as it claims. Four additions do that.

01 · FIXED

Same run every time

One reference set, one seed, one step count, one resolution. Two cards running it are doing identical work, so the clock is the only variable.

02 · SAMPLED

Measured while it runs

The provider app samples NVIDIA DCGM throughout: utilisation, memory in use, clocks, temperature and energy drawn.

03 · SIGNED

A receipt, not a claim

Card, start, end, the samples, the input hash and the output hash are signed by the provider's key. A screenshot cannot fake it.

04 · ANCHORED

Written to the chain

The receipt hash goes on Robinhood Chain. The score on a card's passport points at a block anyone can read.

What a run produces receipt fields

cardModel and the card's hashed identity, never a serial number
started · endedWall-clock bounds of the run, to the second
samplesDCGM series: utilisation, memory, clocks, temperature, watts
input_hashSHA-256 of the fixed reference set and run parameters
output_hashSHA-256 of the produced weights and sample images
provider_sigSignature from the wallet that registered the card
anchorRobinhood Chain transaction carrying the receipt hash

The score is the run's wall-clock time normalised against the reference card, then capped at 100. Faster than reference scores above 50, slower below.

Reference times modelled · fills in from real runs in Phase 2

One TJ Benchmark run costs 0.6 hours whichever card takes it, so a slow card earns less per hour of wall time. That is the point. All cards →

Run it yourself

You do not need a GPU to run the benchmark. Spend hours and a registered card runs it for you, and you get the receipt.

01

Buy hours

1.20 USDG each, paid in your own wallet on Robinhood Chain.

02

Pick the workload

In the Studio, choose TJ Benchmark fine-tune and how many runs.

03

Redeem 0.6 h

The job queues for a verified card. Nothing is spent until it is accepted.

04

Read the receipt

Timing, samples and hashes land in your receipts, and on the card's passport.