GLaM

1.2 trillion parameters

The dataset is made of 1.6 trillion tokens, but later in the paper they say they only train the largest model for 600b tokens. 600b / 0.75 words/token = 800b words. "The complete GLaM training using 600B tokens consumes only 456 MWh and emits 40.2 net tCO2e."

"We use a maximum sequence length of 1024 tokens, and pack each input example to have up to 1 million tokens per batch."

The network activates 96.6 billion parameters per token and trained for 600B tokens. 6 * 600B * 96.6B = 3.478e23 Digitizing figure 4 (d) indicates 139.67 TPU-years of training. 2.75e14 * 139.67 * 365.25 * 24 * 3600 * 0.3 = 3.636e23 Since these are close, we will use the 6NC estimate and derive hardware utilization from the training time information. Later they say they measured 326W power usage per chip, which could maybe be used to estimate utilization.

Note that they give several energy estimates. Use the complete training figures for 600B tokens, not the GPT-3 comparison values with 280B tokens. "326W measured system power per TPU-v4 chip" "The complete GLaM training using 600B tokens consumes only 456 MWh" 1024 TPU v4 chips (456 MWh) / (326W/chip * 1024 chips) = 1366 hours

6ND: 6 * 600B * 96.6B = 3.478e23 Hardware: Digitizing figure 4 (d) indicates 139.67 TPU-years of training. 2.75e14 * 139.67 * 365.25 * 24 * 3600 = 1.212e24 at full utilization Implies utilization of 0.287 MFU = 0.2870

"As shown in Table 5, GLaM (64B/64E) is better than the dense model and outperforms the previous finetuned state-of-the-art (SOTA) on this dataset in the open-domain setting"