VIDRAFT unveils VKORE, software that brings AI training to inference-only NPUs

VIDRAFT has developed VKORE, a software acceleration layer that enables both inference and training on NPUs originally designed for inference only, with no hardware changes. It delivers up to 3.8× faster single-user inference, 4.1× server throughput, and roughly 170× the training throughput of CPU-only baselines. Sixteen patents have been filed, and no public SDK is available yet.

• VKORE boosts inference on the same NPU up to 3.8× and server throughput up to 4.1×
• Transformer training on NPU runs about 170× faster than CPU-only baselines
• The approach is software-only with no silicon changes; 16 patents filed
• Darwin-27B-ZTC-v5 ranked #1 among 152 models on the S1MB benchmark