float16
A numerical precision format that represents a floating-point number in 16 bits (2 bytes). Its storage size is half that of float32, and it has little impact on the retrieval effectiveness of most embeddings.
Timeline (last 90 days)
float16 storage can halve vector size, with little impact on retrieval performance for most embeddings.
Calculated in float16, a probability distribution over a vocabulary of 100K is approximately 200KB.
Calculated in float16, a complete probability distribution for a 100K vocabulary is approximately 200 KB.
All Facts (3)
float16 storage can halve vector size, with little impact on retrieval performance for most embeddings.
50%UnverifiedCalculated in float16, a probability distribution over a vocabulary of 100K is approximately 200KB.
50%UnverifiedCalculated in float16, a complete probability distribution for a 100K vocabulary is approximately 200 KB.
50%