For decades, high-performance computing relied on thirty-two-bit single-precision floating-point arithmetic (FP32) as the gold standard for numerical stability. The deep learning revolution systematically halved this requirement: first to sixteen-bit half precision (FP16 and BF16), and more recently to eight-bit formats (FP8). Yet as model parameters swell into hundreds of billions, the semiconductor industry has set its sights on the next efficiency frontier: four-bit floating-point computation (FP4).
Directly truncating weights and activations to four bits using conventional uniform quantization historically triggered catastrophic accuracy degradation. With only sixteen possible representable values, four-bit representations lack the dynamic range necessary to capture the extreme statistical outliers that inevitably emerge in large transformer layers. Clamping these outlier values introduces massive quantization error, corrupting gradient flow during training and causing severe perplexity spikes during inference.
The Open Compute Project (OCP) Microscaling Data Formats specification, authored cooperatively by NVIDIA, AMD, Arm, Intel, Meta, and Qualcomm, overcomes this barrier through microscopic block scaling. Under the MXFP4 standard, tensors are divided into small, localized vectors of thirty-two elements. Instead of assigning a single scale factor to an entire tensor or row, each thirty-two-element micro-block shares its own dedicated eight-bit scale factor (an E8M0 exponent). The thirty-two individual values within the block are then represented using compact four-bit floating-point values (E2M1).
By localizing dynamic range adjustment to tight thirty-two-element clusters, microscaling isolates outlier values, preventing a single high-magnitude activation from blowing out the quantization grid of adjacent elements. In next-generation accelerator families such as NVIDIA Blackwell, hardware tensor cores execute MXFP4 matrix multiplications natively at twice the throughput and half the energy of FP8. Microscaling proves that precision is not merely a question of total bits, but of structural allocation: placing resolution exactly where information density demands it.