VKORE

VKORE is a software acceleration layer for NPUs, designed to be hardware-agnostic, enabling data centers to mix GPUs and NPUs and still train and serve inference from the same stack.

VKORE aims to rebuild a CUDA-like abstraction layer on NPUs, providing a unified training and inference interface upward and adapting to different vendors' NPU instruction sets downward, so that data centers can swap accelerators like swapping storage or network cards without rewriting model code, thereby breaking single-vendor lock-in.

VKORE is positioned as a hardware-agnostic software layer, aimed at data centers with mixed GPU and NPU deployments, and is not tied to a specific chip vendor.

Currently VKORE and the related pretraining framework are not yet publicly available; the source report did not disclose a Hugging Face repository, GitHub link, or API endpoint.

VKORE's route involves trade-offs: reinventing a cross-hardware portable training instruction set will lead to a substantial rise in software complexity and may lose part of the hardware-level performance optimization space.

The milestones publicly demonstrated by VKORE include: training a 46-million-parameter model from scratch, running on 32 inference-only NPU chips, with a loss correlation of 0.99 with a GPU training baseline.

VKORE works in three layers: computation graph decomposition, gradient orchestration software-ization, and distributed state management, managing gradient flow and accumulation on the NPU array through software-level scheduling.

Based on the currently public evidence, NPUs cannot yet fully replace GPUs for training, because the 0.99 loss correlation holds only for the configuration of 46 million parameters and 32 chips.

As model scale advances toward a billion parameters, memory bandwidth, sustained I/O throughput, and thermal load will become stricter constraints, and inference-only NPUs may not be able to compensate for these physical limits through software abstraction.

VKORE's benchmark test uses Google's open-source Gemma 4 26B model, running on an eight-card, 32-chip Korean NPU server.

When evaluating closed-source acceleration layers such as VKORE, one must be wary of the model weight and adapter portability risks and migration costs that may result from a training path depending on a vendor's proprietary runtime.

In the VKORE scenario, a numerical consistency of 0.9999 usually refers to the correlation coefficient or element-wise match ratio of the output tensors of two computation backends (VKORE and PyTorch) under the same input, not bit-level identity.

VKORE extends the NPU's executable operation set to cover the backpropagation primitives required for LoRA training, and is described as requiring no changes to NPU firmware.

VKORE and PyTorch achieve a numerical consistency of 0.9999 in LoRA training.

The inference acceleration achieved by VKORE is 3.8x for a single user and 4.1x in server throughput.

LoRA fine-tuning is a roadmap item for the VKORE software layer, alongside inference acceleration and pretraining, and there is currently no public timeline.

VIDRAFT verified that general public models can perform both inference and quantized model execution on the same NPU, indicating that the VKORE layer covers both training and deployment ends.

The goal of the VKORE software layer that VIDRAFT is building is to become "the CUDA of the NPU field," covering three capabilities: inference acceleration, full pretraining, and LoRA fine-tuning; what is released this time is a publicly verified artifact of the pretraining milestone.

VIDRAFT is integrating the above capabilities into an NPU software layer called VKORE, and positions it by analogy to the ecosystem advantage Nvidia built through CUDA, describing it as "the CUDA of NPUs".

VIDRAFT's same NPU solution already covers model lifecycle stages including pretraining from scratch, LoRA fine-tuning, inference acceleration (also validated on general open-source models), and quantized model execution and acceleration.