VIDRAFT's VKORE Enables Full Model Pre-Training on Inference-Only NPUs — 0.99 Loss Correlation with GPU Baselines

TL;DR: VIDRAFT Lab has demonstrated that inference-optimized NPUs — hardware traditionally limited to forward-pass workloads — can execute full-scale language model pre-training from scratch using a software abstraction layer called VKORE. A 46M-parameter model trained across 32 inference-dedicated chips achieved a 0.99 loss correlation against standard GPU training runs. For ML engineers, this signals a potential paradigm shift: training infrastructure may no longer require GPU-exclusive hardware stacks.

What it is

VKORE is VIDRAFT Lab's software-defined training layer designed to unlock backpropagation and optimizer state management on NPU hardware that was originally architected for inference-only workloads. The core claim is straightforward but structurally significant: hardware that the industry has historically excluded from the training pipeline can, with the right software orchestration, participate in full model pre-training.

The milestone demonstrated publicly involves:

• A 46M-parameter language model trained entirely from scratch
• Execution across 32 inference-dedicated NPU chips
• A 0.99 loss correlation with equivalent GPU-based training benchmarks

This is not fine-tuning or distillation — it is pre-training, meaning gradient computation, weight updates, and optimizer state accumulation are all happening on hardware that was never designed with that workload in mind.

How it works

At a conceptual level, VKORE functions as a software abstraction layer that remaps the computational graph of a standard training loop — forward pass, loss computation, backward pass, and optimizer step — onto the execution primitives available on inference-optimized silicon.

The high-level approach works roughly like this:

• Compute graph decomposition: The training workload is broken into operations that inference NPUs can natively execute (matrix multiplications, activation functions, elementwise ops), rather than relying on CUDA-specific primitives.
• Gradient orchestration in software: Because inference NPUs lack dedicated hardware pathways for backpropagation, VKORE manages gradient flow and accumulation through software-level scheduling across the chip array.
• Distributed state management: Optimizer states (momentum buffers, variance estimates, etc.) are managed across the 32-chip configuration via software-coordinated memory handling rather than hardware-native all-reduce operations.

The framing from VIDRAFT is explicitly "NPU-as-CUDA" — the goal is to treat the NPU cluster as a general-purpose training substrate, with software compensating for what the hardware was not built to do natively. This is analogous in spirit to how early ML frameworks abstracted over CPU SIMD units before GPU acceleration became standard, but applied in the reverse direction: taking specialized hardware and generalizing it upward through software.

Benchmarks & results

The publicly reported figures are specific:

Metric	Result
Model size	46M parameters
Hardware	32 inference-dedicated NPUs
Training task	Language model pre-training from scratch
Loss correlation vs. GPU baseline	0.99

A 0.99 loss correlation is a strong numerical claim — it means the training trajectory on NPU hardware is nearly statistically indistinguishable from an equivalent GPU run. This is the benchmark VIDRAFT is anchoring the disclosure on, and it is the figure that makes this more than an academic curiosity.

Important caveats noted in coverage: the result is demonstrated at 46M parameters. How this approach scales toward billion-parameter models — where memory bandwidth, sustained I/O throughput, and thermal load become substantially more demanding — has not yet been publicly addressed. Inference-optimized chips are rarely designed for the continuous, high-bandwidth memory access that large-scale training sustains over hours or days.

How to try it

VKORE and the associated pre-training framework have not been announced as publicly available in the source coverage. No Hugging Face repository, GitHub link, or API endpoint has been disclosed at the time of this article.

If and when VIDRAFT makes tooling available, the most likely channels based on their existing public presence would be:

• Hugging Face Hub — for model checkpoints or training configurations
• GitHub — for the VKORE software layer or example training scripts
• OpenAI-compatible API — for inference access to resulting models

Watch VIDRAFT's official channels for announcements. This article will be updated when public access is confirmed.

FAQ

Q: Does this mean NPUs can fully replace GPUs for training today?
A: Not at scale — at least not based on current public evidence. The 0.99 loss correlation result holds for a 46M-parameter model on 32 chips. The open question is whether VKORE's software orchestration approach can sustain comparable fidelity and throughput as model and dataset sizes increase by orders of magnitude. The memory wall and thermal constraints of inference-optimized hardware are real engineering challenges that remain publicly unaddressed at larger scales.

Q: What does "inference-only NPU" actually mean in this context?
A: Inference-optimized NPUs are chips designed to execute a trained model's forward pass efficiently — think low-latency, low-power matrix operations for deployment scenarios. They typically lack hardware support for gradient computation, high-bandwidth bidirectional memory access patterns, and the communication fabric needed for distributed training. VKORE's contribution is making those capabilities available in software, layered on top of what the chip can already do.

---

Originally reported by Mukta AI(글로벌) (2026-10-14) — source article.