Advantages of IF4 Over NVFP4
IF4 reduces quantization error by selecting per‑group between FP4 and scaled INT4 representations, choosing the option with lower mean‑squared error for each group of 16 values.[:cite[1]{ln=1} :cite[2]{ln=1}] IF4 ...
IF4 reduces quantization error by selecting per‑group between FP4 and scaled INT4 representations, choosing the option with lower mean‑squared error for each group of 16 values.[:cite[1]{ln=1} :cite[2]{ln=1}] IF4 achieves lower training loss than NVFP4 during W4A4G4 quantized pre‑training, producing training loss closer to a high‑precision BF16 baseline.[:cite[1]{ln=1} :cite[3]{ln=1}] IF4 yields better post‑training quantization (PTQ) accuracy and lower perplexity than NVFP4 in nearly all evaluated models and sizes.[:cite[1]{ln=1} :cite[4]{ln=1}] IF4 handles inputs transformed by Hadamard (or similarly more uniform distributions) better than NVFP4 by increasing the preference to select INT4 where that yields less error.[:cite[1]{ln=1} :cite[3]{ln=1}] IF4 introduces no memory/storage overhead compared to NVFP4 because it repurposes an unused sign bit in the block scale factor to indicate the chosen representation.[:cite[1]{ln=1} :cite[5]{ln=1}] IF4 can be implemented efficiently in hardware with modest MAC datapath overhead (slightly higher latency and area at the isolated MAC level) while remaining practical because system‑level throughput is often memory‑bound; the paper demonstrates an IF4 MAC design and synthesis results alongside NVFP4 for comparison.[:cite[1]{ln=1} :cite[6]{ln=1}]