Nvidia Launches NVFP4 for Low‑Precision Inference Ahead of GTC 2026 AI Showcase

Nvidia’s Blackwell GPUs now support NVFP4, a 4‑bit floating‑point format that adds a per‑block power‑of‑two scale to the standard E2M1 layout, according to the company’s technical blog. The tweak lets developers quantize models to FP4 without the accuracy loss typical of naïve 4‑bit conversion, while keeping memory footprints minimal.

The blog notes that Blackwell’s fifth‑generation Tensor Cores handle FP4 alongside FP64, FP32/TF32, FP16/BF16, INT8/FP8 and FP6, delivering “efficient, accurate quantization” for inference workloads. A performance chart compares Ampere, Hopper and Blackwell, showing a steep rise in dense and sparse throughput as each architecture adds support for smaller data types.

NVFP4 is positioned as a bridge between the older MXFP4 format— which uses a shared scale per 32‑value block— and the newer FP4 (E2M1) scheme, preserving model fidelity while cutting compute cycles. Nvidia claims the format “builds on the simplicity of earlier formats while maintaining model accuracy,” per the blog’s table of format differences.

The timing aligns with Nvidia’s GTC 2026 AI showcase, where CEO Jensen Huang is expected to demonstrate how the new low‑precision engine tackles token‑heavy generative workloads that have strained the company’s existing GPU line, as reported by The Register. The conference will also highlight Nvidia’s recent acquisition of Groq’s technology, aimed at improving latency‑sensitive token generation.

Analyst firm SemiAnalysis’ InferenceX benchmarks, cited by The Register, show Groq‑derived architectures excelling in the “goldilocks zone” of token throughput. Nvidia’s NVFP4 and Blackwell Tensor Cores are intended to close that gap, offering higher token rates without sacrificing the accuracy needed for enterprise AI applications.