How Firm Is an Answer from a Busy AI Server? Batch Size, Numerical Precision and the Answers of Large Language Models
Large language models are now served from big GPU clusters to millions of people. Recent work shows that the same question can get a different answer on the same model, depending on the batch size, the number of GPUs, the type of GPU and the numerical precision. The usual remedy is to hold these settings fixed so that every answer is computed in the same way. A fixed way of computing does not tell us whether a given answer is firm. This note describes a controller that measures how firm each answer is. It moves a few of the server's own non-learned settings a little, such as batch size, numerical precision and input scaling, with the model and the user's question unchanged. It reads the spread of the answer scores. If the spread is large, it shunts to the next validated setting by itself, without the intervention of a human operator. If no setting is firm, the answer is marked not firm. This is a machine of the second kind: it needs no access to the GPU hardware. The note explains why this is a measurement of behaviour and not Lean proof checking, and why it differs from Bayesian networks and Monte Carlo dropout. It also compares the idea with five papers: numerical sources of nondeterminism in LLM inference, Thinking Machines' work on defeating nondeterminism, AutoEMatic, progressive mixed-precision decoding and TopLoc. Patent pending. Keywords: LLM inference, AI serving, batch size, numerical precision, nondeterminism, firmness of an answer, controller, shunting action, GPU cluster
Authors
- Prakash Vaithyanathan (ORCID: https://orcid.org/0000-0001-8230-4830)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24322151
- Primary Topic
- Natural Language Processing Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00