Unverified50% confidenceOpinionExact time
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.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedThe 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 0.99.78% similarUnverifiedVIDRAFT 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.72% similarUnverifiedThis training ran on a first-generation inference-specific NPU released in 2025, using 32 chips, and is not the latest hardware.72% similarUnverifiedThe number of GPU hardware performance counters is limited, and once dcgm-exporter occupies them, tools such as Nsight Systems and Nsight Compute cannot simultaneously use the same set of counters.70% similarUnverifiedVIDRAFT 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 0.99.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1044259API
curl https://kongchang.com/api/v1/knowledge/claims/1044259MCP
get_claim(id=1044259)