Flopper.io Flopper.io

TF32

precision
Also known as: TensorFloat32, Tensor Float 32

Definition

TensorFloat-32 - A 19-bit precision format introduced with NVIDIA Ampere architecture. Combines the range of FP32 with reduced precision, providing up to 10x speedup for AI training without code changes. Automatically used by Tensor Cores for matrix operations.

GPUs featuring TF32

NVIDIA Rubin

Rubin · 2026 · 288 GB

NVIDIA Vera Rubin

Rubin · 2026 · 576 GB

Huawei Ascend 950DT

Da Vinci v3 · 2026 · 144 GB

NVIDIA B200

Blackwell · 2025 · 180 GB

AMD Instinct MI325X

CDNA 3 · 2025 · 256 GB

NVIDIA B300

Blackwell Ultra · 2025 · 288 GB
Browse All GPUs →

Related Terms

FP32

Single-precision 32-bit floating point, the long-standing default for general-pu...

Tensor Cores

Specialized processing units in NVIDIA GPUs designed to accelerate matrix multip...

BF16

Brain Float 16 - A 16-bit floating-point format developed by Google with the sam...

More precision Terms

BF16

Brain Float 16 - A 16-bit floating-point format developed by Google with the sam...

FP16

16-bit floating-point precision format (half precision). Commonly used in AI tra...

FP8

8-bit floating-point precision format introduced in Hopper architecture. Enables...

Explore More Terms

Browse the full glossary or dive into architecture guides

Back to Glossary Architecture Guides