Quantization

FP8 versus eight-bit integer quantization

What it is

Source: NVIDIA Transformer Engine, E4M3 precision and E5M2 range, read 2026-09-21

Source: vLLM FP8, native compute from CC 8.9 and weight-only from CC 7.5, read 2026-09-21

Source: NVIDIA Turing, integer Tensor Core operations, read 2026-09-21

Why it moves cost and latency

The more precise FP8 format commonly serves weights and activations, while the wider-range format helps represent gradients during training. Hardware support, tensor distributions, calibration burden, and quality targets decide the inference fit.

What it looks like in practice

In vLLM, native FP8 computation is supported on Ada Lovelace, Hopper, and Blackwell GPUs. Turing and Ampere instead use weight-only FP8 with Marlin and higher-precision computation; eight-bit integer Tensor Core support starts with Turing. Teams verify kernel support and quality on the chosen GPU.

Where we measured it

Related terms

Questions this definition answers

What does FP8 versus eight-bit integer quantization mean in inference serving?

FP8 uses an exponent and mantissa within an eight-bit floating-point representation. The four-exponent-bit format favors precision, while the five-exponent-bit format extends range; eight-bit integer quantization instead uses uniformly spaced levels interpreted through scales. In vLLM, native FP8 computation is supported on Ada Lovelace, Hopper, and Blackwell GPUs. Turing and Ampere instead use weight-only FP8 with Marlin and higher-precision computation; eight-bit integer Tensor Core support starts with Turing. Teams verify kernel support and quality on the chosen GPU.