PulseAugur
EN
LIVE 09:22:33
한국어(KO) 추론 전용 NPU로 LLM 사전학습이 가능할까? 3억 2,500만 파라미터를 0부터 학습해 B200과 같은 품질을 냈다

Inference-only NPU matches GPU quality in LLM pre-training

By PulseAugur Editorial · [1 sources] ·

Vidraft Lab has successfully pre-trained a 325 million parameter language model, VKORE-325M-NPU-Base, from scratch using the REBELLION ATOM-Max, a commercial inference-only NPU. This achievement challenges the notion that inference-focused NPUs cannot be used for training. The NPU's performance was comparable to that of an Nvidia B200 GPU when evaluated on unseen text data, with a loss of 0.7993 versus 0.7994 respectively. Vidraft has open-sourced the model weights and evaluation scripts, though their proprietary training framework and stabilization techniques, crucial for the successful training, remain under patent. AI

IMPACT Demonstrates that inference-focused NPUs can be viable for LLM pre-training, potentially reducing hardware costs and increasing accessibility.

RANK_REASON Research paper detailing successful LLM pre-training on an inference-only NPU. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

Inference-only NPU matches GPU quality in LLM pre-training

How we ranked this

Signal score
20 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
Research paper detailing successful LLM pre-training on an inference-only NPU. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
model release, infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. dev.to — LLM tag TIER_1 한국어(KO) · Vidraft_lab ·

    Can LLMs be pre-trained with inference-only NPUs? We trained 325 million parameters from scratch and achieved quality comparable to B200.

    <h2> 요약 </h2> <p>가능합니다. 단, 학습용 소프트웨어를 따로 만들어야 합니다. 비드래프트는 3억 2,560만 파라미터 언어모델 <strong>VKORE-325M-NPU-Base</strong>를 무작위 초기값에서 출발해 <strong>리벨리온 ATOM-Max</strong>(2025년 출시된 상용 추론 전용 NPU)로 사전학습했습니다. 같은 모델을 <strong>B200 GPU</strong>로도 학습해 비교했습니다. 두 모델 모두 학습에 한 번도 쓰지 않은 텍스트 1,024개 구간에…

RELATED ENTITIES

RELATED TOPICS