Unverified50% confidenceBenchmarkExact time
The NPU server's training throughput is about 54% of a single NVIDIA B200, while the hardware price is comparable.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/13/2026
First Seen
Valid until: 10/27/2026
Sources
Related Entities
Related Claims
UnverifiedM2 Ultra满载功耗约60-80W，而同等推理能力的NVIDIA服务器组合功耗动辄超过1000W63% similarUnverifiedRunning a 70B model locally requires at minimum a professional GPU with 48GB of VRAM, such as an NVIDIA A6000 or dual RTX 4090 cards.61% similarUnverifiedUsing NVIDIA B300 chips in a 1-megawatt data center can achieve a total throughput of approximately 450,000 tokens/s, which at 340 tokens/s per user amounts to approximately 1,324 concurrent active users.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013621API
curl https://kongchang.com/api/v1/knowledge/claims/1013621MCP
get_claim(id=1013621)