MXInt-based datapath designs put all ViT nonlinear operators on an FPGA at 2-5 bit mantissas, with under 1% ImageNet accuracy loss on DeiT models.
Revisiting block-based quantisation: What is important for sub-8-bit llm inference?,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.AR 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Refining Datapath for Microscaling ViTs
MXInt-based datapath designs put all ViT nonlinear operators on an FPGA at 2-5 bit mantissas, with under 1% ImageNet accuracy loss on DeiT models.