Expand your power
On the ROCK 5C 8GB, Spectrum replaced 9 out of 20 steps with state prediction. Total image generation time decreased by 41.02%, and speed and image quality were compared against…
FFN W8A8 quantization reduced denoising time by 18.48%, but failed to meet image quality criteria. Model residency also increased swapping at 18 contexts. We explore why neither…
Building the AdaLN graph that crashed in RKNN Toolkit 2.3.2 using 2.3.0 enabled fusing a single DiT block. The speedup was limited to 0.47%, while NPU device mapping memory…
We split Anima into 84 FP16 RKNN models and ran it on a ROCK 5C 8GB. This post examines CPU and NPU roles, pre- and post-conversion output comparison, and latency for denoising and…