BLOOM's footprint, beyond the current
On 3 November 2022 a full life-cycle account appeared for BLOOM, an open 176-billion-parameter model. Training on 384 A100 accelerators at the French IDRIS centre ran for 118 days and drew 433,196 kWh. The current accounted for 24.69 tonnes of CO2 equivalent, the manufacture of the hardware for 11.2, and the cluster's idle time for 14.6; 50.5 tonnes in all.
Why it matters
Earlier calculations stopped at the electricity consumed. Here three terms are set side by side for the first time - current, hardware manufacture and idle time - and the current alone turns out to be under half of it. After this, an account that reads only the meter is visible as partial rather than complete.
The shares: dynamic consumption 48.9 per cent, embodied emissions 22.2 per cent, idle 28.9 per cent. The French grid's carbon intensity is 57 grams of CO2 equivalent per kilowatt-hour, about seven and a half times lower than the data centre in the previous record. Together with a tenfold smaller draw of current, that is what separates 552 tonnes from 25. The run lasted 118 days, 5 hours and 41 minutes and took 1,082,990 accelerator hours - an average of 384 of them across 48 nodes. What the record does not claim. The embodied emissions are an estimate, not a measurement: the authors took a published 2,500 kg footprint for the HPE server and an outside estimate of about 150 kg for an accelerator, because NVIDIA publishes no figure of its own. The paper calls itself the first attempt to count the broader footprint of a large language model; that is its own wording.