#NPU
4 related articles
Training Large Models Without GPUs: Backpropagation Runs on Inference-Only NPUs
VIDRAFT pretrained a 46M-parameter language model VKORE-46M-NPU from scratch on an inference-only NPU, achieving a 0.99 loss trajectory correlation with GPU training. This article analyzes its engineering methods, technical limitations, and practical significance.
VKORE: Making Inference-Only NPUs Run AI Training
VIDRAFT launches VKORE, a hardware-agnostic software layer letting inference-only NPUs support AI training. 3.8× inference speedup, ~170× CPU training throughput, 16 patents filed.
NPUs Can Pretrain Too: VKORE Hits 170% of H100 Throughput on Domestic Chips
VKORE software acceleration layer enables NPUs to support pretraining, with measured throughput reaching 170% of a single H100 and 75% of a B200, plus 3.8x inference gains—offering a full training loop for sovereign compute without Nvidia.
Inference-Only NPUs Can Now Train Transformers: VKORE Achieves 0.9999 Precision Alignment
VIDRAFT's VKORE software layer enables inference-only NPUs to achieve multi-fold inference speedups and LoRA training, with 0.9999 precision alignment to PyTorch.