Rongchai Wang
Feb 23, 2026 18:39
NVIDIA’s NVFP4 4-bit coaching format achieves 59% quicker AI mannequin coaching than BF16 whereas matching accuracy on Llama 3 8B benchmarks, per new analysis.
NVIDIA’s NVFP4 low-precision coaching format delivers as much as 1.59x quicker throughput in comparison with normal BF16 coaching whereas sustaining equal mannequin accuracy, in response to new benchmarks revealed by the corporate’s analysis workforce on February 23, 2026.
The outcomes mark a major milestone for 4-bit AI coaching, demonstrating that aggressive numerical compression would not require sacrificing mannequin high quality when correct methods are utilized.
The Numbers That Matter
Testing on Llama 3 8B fashions skilled throughout 1 trillion tokens, NVIDIA’s workforce measured throughput at 1,850 TFLOP/s per GPU with NVFP4 versus 1,165 TFLOP/s for BF16 baseline—a 59% enchancment. The checks ran on GB200 NVL72 {hardware} utilizing the corporate’s Blackwell structure.
Downstream benchmark scores inform the actual story. On MMLU, NVFP4-trained Llama 3 8B scored 45.64% in comparison with 45.98% for BF16. HellaSwag confirmed 75.59% versus 76.44%. These variations fall inside noise margins for sensible purposes.
Reminiscence effectivity positive aspects enabled doubling the micro-batch dimension from 2 to 4 throughout pretraining, instantly enhancing scalability for large-scale coaching runs.
Why 4-Bit Coaching Works Now
Earlier makes an attempt at ultra-low-precision coaching usually resulted in mannequin divergence or vital accuracy degradation. NVIDIA’s method sidesteps these points via a particular recipe that is emerged from intensive testing.
The vital perception: holding roughly 15% of the community in larger precision prevents coaching collapse. Particularly, the ultimate 4 transformer layers should stay in BF16. Ablation research confirmed that totally NVFP4 fashions diverge throughout coaching.
The format makes use of a two-level scaling technique—micro-block scaling for teams of 16 components mixed with international FP32 scaling throughout full tensors. This hierarchical method manages the restricted dynamic vary inherent in 4-bit representations.
Random Hadamard transforms easy tensor spectrums and cut back outliers that will in any other case trigger coaching instability. Stochastic rounding for gradients eliminates systematic quantization bias.
Comparability With Different Low-Precision Codecs
NVFP4 is not the one choice. FP8 with present scaling (FP8-CS) achieved 1.33x speedup over BF16, whereas MXFP8—a block-level scaling variant optimized for Blackwell—hit 1.32x. Each codecs confirmed barely higher convergence monitoring than NVFP4 throughout coaching, although ultimate accuracy metrics remained comparable throughout all approaches.
MXFP8 demonstrated marginally higher efficiency than normal FP8, seemingly as a consequence of finer-grained scaling that higher captures native dynamic vary inside tensors.
Manufacturing Deployment
The methods can be found now via NeMo Megatron Bridge, NVIDIA’s open PyTorch-native library. Switching between precision codecs requires altering a single configuration flag—no mannequin code or optimizer logic modifications wanted.
For groups operating large-scale coaching workloads on Blackwell {hardware}, the throughput positive aspects translate on to decreased coaching time and compute prices. A mannequin that beforehand required 10 days of coaching may doubtlessly full in below 7 days with NVFP4.
The really useful recipe for NVFP4: AdamW optimizer with epsilon=1e-8, studying charge decaying from 6e-4 to 6e-6, and international batch dimension of 768. These parameters symbolize the empirical candy spot from NVIDIA’s intensive testing throughout a number of architectures and datasets.
Picture supply: Shutterstock
