A single software layer could eliminate the need for separate GPU infrastructure in AI data centres.
TL;DR: VIDRAFT, a Korean AI startup, has demonstrated that its VKORE software acceleration layer can run LoRA fine-tuning of transformer models on NPUs normally reserved for inference. VKORE achieves 0.9999 numerical parity with standard PyTorch during training and accelerates single-user NPU inference by 3.8× while lifting server throughput by 4.1×. The same NPU training workload runs approximately 170× faster than an equivalent CPU on the same server.
VIDRAFT, the Korean Pre-AGI AI startup, made waves in the AI infrastructure world on October 14, 2026, when global tech outlet TechGig reported that the company's VKORE software layer had successfully enabled LoRA training on NPUs that are conventionally restricted to inference workloads — a capability that could meaningfully reshape how data centres allocate and spend on AI hardware.
At the heart of VIDRAFT's announcement is VKORE, a hardware-independent software acceleration layer designed to sit between AI workloads and the underlying silicon — whether that silicon is a GPU or an NPU. The key breakthrough is that VKORE allows LoRA (Low-Rank Adaptation) fine-tuning of transformer models to run on NPUs that have traditionally been limited to inference-only tasks.
In practical terms, VKORE demonstrated two distinct performance benefits on NPU hardware. First, it boosts single-user inference latency by 3.8× compared to running without the layer. Second, it improves server-level throughput — the rate at which a machine handles many concurrent requests — by 4.1×. Both gains are delivered on the same NPU hardware, with no additional chips required.
For the training side, VIDRAFT's results showed that NPU-based LoRA training with VKORE runs approximately 170× faster than training on the CPU present in the same server — a striking gap that underlines how underutilised NPU compute has been for training tasks. Critically, the numerical output of VKORE's LoRA training achieved 0.9999 parity with PyTorch, the de facto standard deep learning framework, meaning model quality is effectively preserved.
Measurements were carried out using a Korean NPU server equipped with eight cards and 32 chips, running Google's open-weight Gemma 4 26B model. VKORE is engineered to be hardware-independent, making it applicable across data centres that deploy a mix of GPUs and NPUs.
The conventional wisdom in AI infrastructure has long held that training and inference are fundamentally separate concerns requiring separate hardware. GPUs, with their flexible compute profiles and mature training ecosystems, handle model training and fine-tuning. NPUs, optimised for the narrower, high-throughput demands of serving models to end users, handle inference. This division has driven enormous capital expenditure: organisations that want to fine-tune models must maintain or rent GPU clusters on top of their existing NPU inference fleets.
VIDRAFT's VKORE directly challenges that orthodoxy. If an NPU inference fleet can perform LoRA fine-tuning during low-traffic periods — overnight, during off-peak hours, or between demand spikes — organisations could potentially eliminate a separate GPU fine-tuning infrastructure entirely. The economics here are significant: inference NPU fleets are already paid for and sitting at partial utilisation during quiet periods. Turning that idle compute into a fine-tuning resource at 0.9999 PyTorch parity is a compelling proposition for any cost-conscious AI operator.
Beyond cost, there is a strategic dimension. As purpose-built NPU deployments expand — particularly in markets like South Korea, where domestic AI chip development is a national priority — the ability to unlock training capabilities on those chips without replacing or augmenting them with GPUs could accelerate NPU adoption globally. VKORE positions VIDRAFT as a key enabler of that transition, sitting above the hardware layer and abstracting the complexity away from the end user.
Q: What is VKORE and who makes it?
A: VKORE is a hardware-independent software acceleration layer developed by VIDRAFT, a Korean Pre-AGI AI startup. It is designed to enhance AI workload performance on both GPUs and NPUs in data centre environments.
Q: How accurate is VKORE's LoRA training compared to standard PyTorch?
A: VKORE achieves 0.9999 numerical parity with PyTorch during LoRA training of transformer models, meaning its outputs are effectively identical to those produced by the standard deep learning framework.
Q: Can existing NPU inference servers run fine-tuning with VKORE without hardware changes?
A: Yes. According to TechGig's reporting, VKORE enables LoRA fine-tuning on NPUs already deployed for inference, with no additional hardware required — allowing operators to use idle capacity during low-traffic periods for model fine-tuning tasks.
Source: TechGig(글로벌) (2026-10-14) — original article