trends.barnas.net

hardware · first seen 4 Oct 2026

bfloat16

Also: bf16, FP32 upcasting, mixed-precision inference, bfloat16 mixed-precision training

A 16-bit floating-point format with fp32 exponent range used to speed up neural network computation.

Mentions per week

0219 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Earliest sources

Latest sources

Mentioned together with

mixture of expertscross-vendor numerical mismatchbatch-invariant inferencequantizationvision-language modellow-precision floating-point arithmeticKV cache compressionstochastic roundingmodel weight compression