… IF4 (Int/Float 4) data type selects between FP4 and INT4 … The selected data type is denoted
using the scale factor’s … outperforms existing 4-bit block-scaled formats, achieving lower loss …

… block-scaled datatypes and implement GeMMs using them on ARM vector processor. • We
develop an adaptive block-scaled … “fine-adaptation” to allow accuracytraining time trade-off. …

… operations of our block-scaled kernels. We also provide more ablation studies for the
adaptive training setup to justify our approach. Our main contributions are as follows: …

… vector shapes for blockscaled datatypes and implement … adaptive block-scaled DNN
training approach, by initially performing “coarse-commutative” training followed by “fine-adaptation…

… Table 1 gives very diverse examples of such data objects. In this context we refer to each
dataset as a data type to indicate that it comes from a distinct mode of measurement or domain. …

… While Four Over Six could theoretically be applied to other block-scaled low-precision FP4
… factors: a block scaled to 4 requires a scale factor that is 50% larger than a block scaled to 6. …

… Additionally, model adaptation at the edge is often required to effectively handle data drift, …
Block-scaled GeMM kernels are designed and optimized across multiple design parameters…

… We focus on convolutions of block– scaled 16 bit data both … chain using fixed–point data
types, we based our software … , and the block exponent is adapted appropriately. This results in a …

… quantization with support for block-scaled data formats. Our … variety of number formats,
including the block-scaled ones, to aid the … Our experiments confirm that block-scaled data formats …

… For this edge training, Microscaling (MX) data types offer a … unit that supports all six MX
data types by exploiting sub-word … [33] NS Murthy et al., “Adaptive block-scaled gemms on …