> "For the largest model size of 13B parameters, the MatMul-free LM uses only 4.19 GB of GPU memory and has a latency of 695.48 ms, whereas Transformer++ requir

> "For the largest model size of 13B parameters, the MatMul-free LM uses only 4.19 GB of GPU memory and has a latency of 695.48 ms, whereas Transformer++ requires 48.50 GB of memory and exhibits a latency of 3183.10 ms" That's a really, _really_ big difference in memory usage and since this scales sub-linear (300M param model uses 0.21GB, 13B model uses 4.19B) a 70B model would fit on an RTX 4090. I think currently people often run 34B Models with 4bit quants on that so I would like to see some larger models trained on more tokens with this approach. Also their 2.7B Model took 173hours on 8 NVIDIA H100 GPUs and that also seems to roughly scale linearly with the parameter size, so a company with access to a small cluster of those DGX pods (say 8) could train such a model in about 30 days - though the 100B token training set might be lackluster for SotA but maybe someone else could chime in on that.