Unverified50% confidenceBenchmarkExact time
An NPU equipped with VKORE reaches 170% of a single H100's per-device pretraining throughput.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/13/2026
First Seen
Valid until: 10/27/2026
Sources
Related Entities
Related Claims
UnverifiedThe speed of LoRA training on the NPU is about 170 times that of the CPU of the same server.71% similarUnverifiedOn an A100 (80 GB VRAM) GPU, the throughput for training a 7B model is approximately 3,000–5,000 tokens per second.60% similarUnverifiedThe mxfp scheme (stochastic rounding + Hadamard preprocessing) achieves 37.2K tokens/s/GPU, corresponding to 86.3% MFU, with a final training loss 2.11% higher than pure bfloat16.59% similarUnverifiedFor small teams with a few professional-grade GPUs (such as A100 80GB or H100), a 270GB model has entered the range of operability.58% similarUnverifiedThe TEMPEST model has only 720,000 parameters, a storage footprint of 2.80 MB, and is able to converge within 50 training epochs.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1017226API
curl https://kongchang.com/api/v1/knowledge/claims/1017226MCP
get_claim(id=1017226)