convergence-study-hardware-bfloat16-greedy-decoding

IN premise — summaries/2026/08/24/convergence-without-understanding-2026-sR-references.md

Created 2026-08-24T17:10:53+00:00

The convergence study used a single NVIDIA RTX 5090 (32 GB VRAM) for most models, 2× A100 80 GB for 70B models, bfloat16 precision, and greedy decoding (temperature = 0).

Summary

This locks in the exact hardware and software setup behind every convergence measurement, so anyone trying to reproduce or challenge the results knows the precise baseline they are working from. The combination of reduced-precision arithmetic and zero-temperature decoding also means the outputs are deterministic and the numbers are sensitive to that specific precision choice, which matters when comparing results across studies.