Knowledge Base / Glossary

Dtype

The data type (dtype) of a tensor or array: the storage format shared by every element, such as float32, bfloat16, or int8. A dtype label names storage, not the whole arithmetic path.

A dtype (data type) is the storage format shared by every element of a tensor or array: how many bits each element occupies and how those bits are interpreted as a number. In PyTorch and NumPy the dtype is a property of the whole tensor, not of individual elements.

Dtype	Bytes per element	Interpretation
`float32` (FP32)	4	32-bit floating point
`float16` (FP16)	2	16-bit floating point (IEEE 754 binary16)
`bfloat16` (BF16)	2	16-bit brain floating point
`int64`	8	64-bit signed integer (default for indices)
`int8`	1	8-bit signed integer (common in quantized inference)
`bool`	1	true/false mask element

import torch

t = torch.tensor([1.0, 2.0])       # default floating dtype: float32
print(t.dtype)                     # torch.float32
h = t.to(torch.bfloat16)           # cast: convert to another dtype
print(h.dtype, h.element_size())   # torch.bfloat16 2 (bytes per element)

Casting converts a tensor to another dtype, rounding each element onto the target format’s representable grid. Casting to a shorter floating dtype can change values; see BF16, FP16, and FP32 for how the floating-point dtypes trade range against precision.

What a dtype label does not establish

A dtype names storage, not the full arithmetic path. In particular:

It does not name the multiplication or accumulation precision. GPU matmul kernels commonly multiply FP16 or BF16 operands while accumulating in FP32.
It does not name the kernel or hardware path (Tensor Cores versus general-purpose lanes, TF32 rounding of FP32 inputs).
It does not make results bit-reproducible or comparable across devices or library versions.

The Deep Dive BF16, FP16, and FP32: Precision, Range, Swamping, and Determinism works through these distinctions with runnable demos and receipts.