Unverified50% confidenceCautionExact time
EmbeddingGemma 2 inference must use BF16 or FP32 and should not use FP16, because activation values exceed the dynamic range of FP16, which can cause NaN values or quality degradation without any error being raised.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe maximum value of FP16 is approximately 65504; its dynamic range is smaller than that of BF16 and is prone to overflow issues.70% similarUnverifiedThe conversion from FP8 to BF16 is lossless because BF16 can exactly represent all values representable by FP8 E4M3, while the reverse is not true67% similarUnverifiedCompared to FP16 original precision, FP8 quantization typically causes no more than 1–2% performance degradation on mainstream benchmarks.66% similarUnverifiedBF16采用1位符号位、8位指数位和7位尾数位，牺牲精度但保留了与FP32相同的动态范围65% similarUnverifiedINT4量化将原本由65536个不同值（FP16）表示的权重压缩到仅16个离散级别65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/985567API
curl https://kongchang.com/api/v1/knowledge/claims/985567MCP
get_claim(id=985567)