# What Does 4-Bit Microscaling Cost a Wearable Activity Recognizer?

DevFeed: [What Does 4-Bit Microscaling Cost a Wearable Activity Recognizer?](<https://devfeed.tech/articles/what-does-4-bit-microscaling-cost-a-wearable-activity-recognizer-116912.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.04185>)

Author: Megan Healy, Hossein Valavi

Published: 2026-11-04T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [quantization](<https://devfeed.tech/topics/quantization.md>), [NVFP4](<https://devfeed.tech/topics/nvfp4.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>)

Tags: [2023](<https://devfeed.tech/tags/2023.md>), [2025](<https://devfeed.tech/tags/2025.md>), [ablation](<https://devfeed.tech/tags/ablation.md>), [accelerometer](<https://devfeed.tech/tags/accelerometer.md>), [actions](<https://devfeed.tech/tags/actions.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [basketball](<https://devfeed.tech/tags/basketball.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [bits](<https://devfeed.tech/tags/bits.md>), [block](<https://devfeed.tech/tags/block.md>), [blocks](<https://devfeed.tech/tags/blocks.md>), [clip](<https://devfeed.tech/tags/clip.md>), [cnn](<https://devfeed.tech/tags/cnn.md>), [comparison](<https://devfeed.tech/tags/comparison.md>), [conversion](<https://devfeed.tech/tags/conversion.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cross-validation](<https://devfeed.tech/tags/cross-validation.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data](<https://devfeed.tech/tags/data.md>), [deep-learning](<https://devfeed.tech/tags/deep-learning.md>), [wearable](<https://devfeed.tech/tags/wearable.md>)

## AI overview

In a 159,492-parameter CNN recognizing four basketball actions from wrist-accelerometer data, simulated 4-bit quantization preserved more macro-F1 with NVFP4 than MXFP4, especially when both weights and activations were quantized: losses were 0.75 and 5.14 percentage points, respectively. NVFP4 led in all five held-out-player folds and under two alternative simulator settings. The ablation attributes much of MXFP4's loss to its simulated round-down scale selection, which can make scales too small and clip large values. Rounding scales up recovered 3.18-3.63 points; using an 8-bit floating-point scale added 0.84-1.20 more. Block size had less consistent impact. Measured TFLite FP16 and INT8 lost at most 0.16 points. Results are from the supplied abstract and simulated 4-bit evaluations.

## Source excerpt

arXiv:2611.04185v1 Announce Type: new Abstract: Wearable sports sensors can recognize actions on the device itself, where lower-precision numbers reduce model storage and memory traffic. Two 4-bit block-scaled formats, MXFP4 from the Open Compute Project, introduced in 2023, and NVFP4 from NVIDIA, introduced in 2025, are drawing interest. Published evaluations cover general deep-learning benchmarks and large language models; to our knowledge, nobody has yet measured what they cost a small wearable-sensor model. We train a 159,492-parameter 1D CNN to recognize four basketball actions from wrist-accelerometer data in Hang-Time HAR and test it on players it has never seen, using five-fold cross-validation grouped by player. Every comparison between number formats is made fold by fold. In full precision the model reaches 69.26\% macro-F1. Measured TFLite FP16 and INT8 models lose at most 0.16 percentage points. In simulation, 4-bit weights cost 1.45 points with MXFP4 and 0.72 points with NVFP4. When activations are quantized to 4 bits as well, MXFP4 loses 5.14 points and NVFP4 only 0.75, and NVFP4 comes out ahead in all five held-out-player folds. The ranking survives two alternative simulator settings, with NVFP4 leading by 3.49 to 4.40 points. A $2 \times 3$ ablation with 4-bit weights and activations (two block sizes and three scale rules) points to scale selection as the main source of MXFP4's loss. The reference MXFP4 conversion we simulate rounds the scale down, which can pick scales that are too small, so many blocks clip their largest values. Rounding the scale up recovers 3.18 points with 32-value blocks and 3.63 points with 16-value blocks, and switching to an 8-bit floating-point scale, as NVFP4 does, adds another 0.84 and 1.20 points. Block size matters less. Under this round-down rule, 16-value blocks score 0.44 points below 32-value blocks in all five folds; under the other two rules there is no consistent effect.