Contribute

Every Diescope result here comes from a report folder you can rerun. There are two ways to add to it.

Diescope isn't released yet. Until it is, write to [email protected] and we'll send a build.

1. Run the standard suite on your hardware

The suite (bench v1) is a fixed, versioned set of workloads: a CNN and a transformer in training, and llama.cpp prompt and generate. Every machine runs the same commands, so results compare one variable at a time.

Set up the workloads (the only step that uses the network; about 12 GB on disk, including a 4.7 GB model)

diescope bench install

Run them, five rounds by default

diescope bench run

It needs an NVIDIA GPU and the CUDA toolkit for the llama.cpp workloads. Other chips come later.
bench run refuses battery power unless given --allow-battery, and a virtual machine unless given --allow-vm, and labels those results. It refuses when other processes use the GPU.
It writes a folder named bench-v1- and the time. Send the whole folder.
Reports leave out host names, user names, and serial numbers. Home folders read as ~.

2. Submit a workload

A workload you'd like measured across hardware: a model, a training job, an inference server, a video pipeline. Send:

The exact command, with versions of everything it runs.
Where its inputs come from. Public data and models only.
Training or inference, and how long one run takes.
Why it matters: the hardware and the question you want answered.

Accepted workloads join a future suite version, so earlier results stay comparable.