Adaptive Block-Scaled Data Types

Dimensions overall score 9.0

• IF4 outperforms existing 4-bit block-scaled formats in quantized training and post-training quantization tasks.result 90%
• NVFP4 suffers from large quantization error on near-maximal values in each group of 16 values due to its error distribution.limitation 90%
• IF4 can be implemented efficiently in next-generation hardware accelerators with a designed Multiply-Accumulate (MAC) unit.technical 85%
• IF4 selects between FP4 and INT4 representations for each group of 16 values using the scale factor's sign bit, which is unused in NVFP4.method 95%
• Adaptive Block-Scaled Data Types provide better model performance for most low-precision memory constraints compared to other formats.result 80%
• The success of IF4 heavily depends on hardware support and adoption by the broader machine learning community, with potential integration issues slowing adoption.market 75%
• IF4 results in less quantization error across many weight channels in large models like Qwen3.5-35B-A3B with no storage overhead compared to NVFP4.result 85%
• The same insight used for IF4 can be applied to design formats for other bit-widths, including IF3 and IF6.method 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card