Bdraft (비드래프트) unveils 'VKORE,' an acceleration software layer for running both inference and training on NPUs (report)

Verified (fact)

Bdraft unveiled 'VKORE,' an acceleration software layer that improves NPU inference performance and also enables training. It works like NVIDIA's CUDA for GPUs, raising compute efficiency on NPUs, and the company said it delivers performance through software alone without changing chips. In a LoRA training test on a transformer model, a job that took one hour on a single B200 GPU took about 320 hours on a server CPU, and running it on an NPU in the default way took even longer than the CPU.

Needs fact-check (claim)

NPU training performance with VKORE applied is cut off in the excerpt and cannot be confirmed.

Why it matters (interpretation)

It targets the weak software ecosystem and lack of training support that hold back the use of Korean-made NPUs.

Sources

• Press 비드래프트, NPU 가속 소프트웨어 ‘VKORE’ 개발… 추론·학습 동시 구현 AI Times (Korea)