Search papers, labs, and topics across Lattice.
The University of Hong Kong (HKU), Hong Kong
6
0
6
4
ScaleLUT is presented, a hardware-oriented LUT design framework and fully parallel reconfigurable accelerator for real-time multi-scale SR deployment and demonstrates the effectiveness of joint LUT algorithm-hardware co-design for practical and energy-efficient edge SR deployment.
Early denoising stages in diffusion models are up to 2.58脳 more vulnerable to hardware noise, but a novel approach can mitigate this without retraining.
Recalibrating the CFG scale in response to CIM noise can restore over 87% of generation quality lost due to nonidealities in Diffusion Transformers.
Hybrid-LUT slashes LUT storage by two-thirds while enhancing image denoising quality, outperforming existing methods by a significant margin.
Accepting selective mismatches in autoregressive decoding can boost throughput by over 15% without any additional training or model adjustments.
ReTopK achieves a 3.07x speedup in attention computation with only a 0.50% increase in perplexity, revolutionizing long-context processing efficiency.
