Unverified50% confidenceEventExact time
VIDRAFT Lab announced an experiment: an NPU array designed solely for inference completed full pretraining of a language model from scratch under the scheduling of the VKORE software layer, and the correlation of its loss curve with a GPU baseline reached 0.99.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/16/2026
First Seen
Valid until: 10/30/2026
Sources
Related Entities
Related Claims
UnverifiedVIDRAFT pre-trained a 46-million-parameter language model, VKORE-46M-NPU, from scratch on a class of NPU hardware normally used only for inference, with the entire training process not relying on NVIDIA GPUs or the CUDA software stack.78% similarUnverifiedThe experiment ran the same model, same initialization, same data and same training schedule on a 32-chip NPU cluster and a reference GPU, and the correlation between the NPU and GPU loss trajectories reached 0.99.75% similarUnverifiedOn a GPU-less Debian machine, using OpenCode 1.19.0 connected to qwen3:1.7b in Ollama for pure CPU inference, three tests all failed to produce acceptable graphics: the token bucket rate limiter task ran for 16 minutes without writing any file; the CDN edge cache task wrote a 738-byte invalid SVG in 8 minutes 23 seconds; and the database connection pool task output a flat 2D placeholder graphic in 10 minutes 37 seconds.70% similarUnverifiedBased 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.69% similarUnverifiedVIDRAFT's VKORE software acceleration layer, without changing the chip, both increases NPU inference performance several times over and makes Transformer LoRA training run on this kind of inference-specific hardware, with numerical precision aligned with standard PyTorch at 0.9999.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1044251API
curl https://kongchang.com/api/v1/knowledge/claims/1044251MCP
get_claim(id=1044251)