MicroQonv: Reshaping Convolution Tensors for Efficient Microscaling in Training and Inference
MicroQonv cuts the hidden memory and requantization costs that make microscaling awkward for convolution layers.
The paper says naive microscaling can quantize tensors twice and move more full-precision data than expected. Its method quantizes each tensor once, then applies a channel-batch-first version of im2col to keep activation overhead down. The authors report a 2x quantization-cost reduction for weights and gradients, up to 9x for activations, and up to 7.53x lower memory movement and storage versus full precision. They also report lower activation traffic on YOLO nano models and accuracy gains for 4-bit continual learning at the edge. ArXiv · AI/CL/LG's note
The paper says naive microscaling can quantize tensors twice and move more full-precision data than expected. Its method quantizes each tensor once, then applies a channel-batch-first version of im2col to keep activation overhead down. The authors report a 2x quantization-cost reduction for weights and gradients, up to 9x for activations, and up to 7.53x lower memory movement and storage versus full precision. They also report lower activation traffic on YOLO nano models and accuracy gains for 4-bit continual learning at the edge. ArXiv · AI/CL/LG's note
score 4