A post on the Hugging Face Blog reports that the same compute cluster achieved a 33-point increase in utilization. According to the title, the improvement came not from changing the hardware but from changing the order in which work was arranged.
No further details from the article body were provided in the input, so the specific methods, workloads, or measurement approach behind the reported figure are not described here.
Why it matters
Utilization gains on existing hardware can affect the cost and efficiency of running compute-intensive workloads without additional infrastructure investment. The reported result suggests scheduling or ordering choices can materially influence how fully a cluster is used.