Rat metrics Explained!

What is a Rat Score?

Rat Scores are computed every 15 minutes to reflect recent GPU usage behavior. The formula for computing the score at each update interval is:

\[ \text{RatScore} = (\text{GPUs})^{1.5} \cdot \log(\text{runtime}_{\text{days}} + 1) \]

Where:

GPUs is the number of GPUs currently in use by the user.
runtime is the duration in days for which the job has been running.

What is the Rat Index?

The Rat Index is a smoothed, time-aware metric that captures historical GPU usage behavior over time. It is computed using an Exponential Moving Average (EMA) over the Rat Scores, with a half-life of 7 days.

Mathematically, for each user, the Rat Index \( R(t) \) at time \( t \) is updated as follows:

\[ R(t) = R(t_0) \cdot e^{- \frac{\Delta t}{\tau}} + S(t) \cdot \left(1 - e^{- \frac{\Delta t}{\tau}} \right) \]

Where:

\( R(t_0) \): the previous Rat Index value
\( S(t) \): the current Rat Score (if available)
\( \Delta t \): time elapsed (in minutes) since the last update
\( \tau \): the EMA time constant, defined as:

\[ \tau = \frac{\text{half-life} \cdot 24 \cdot 60}{\ln 2} \]

If a user has no new activity, their Rat Index decays exponentially over time.

New users are initialized directly with their first recorded Rat Score:

\[ R(t_0) = S(t_0) \]

🐀🐀🐀 Rat metrics Explained! 🐀🐀🐀

What is a Rat Score?

What is the Rat Index?