Unverified60% confidenceBenchmarkExact time
在700亿参数DeepSeek R1蒸馏模型上，因模型溢出RTX 5090的32GB显存，M5 Ultra比5090快20倍以上
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
9/23/2026
First Seen
Valid until: 12/22/2026
Sources
Related Entities
Related Claims
Unverified700亿参数负载下5090约275瓦、M5 Ultra约300瓦，功耗相近但M5 Ultra快20倍以上76% similarUnverified对照模型Supra-50M有51.8M参数，bf16下103.6MB，8-bit量化后56.2MB，降到int4困惑度从165恶化到193，压到三元直接崩溃69% similarVerified700亿参数（70B）的模型经量化后可以在单张RTX 4090（显存24GB）上运行69% similarUnverifiedMixtral 8x7B拥有约467亿总参数，但每次前向传播仅激活约130亿参数，推理计算量接近13B模型，仍需约90GB显存（FP16）68% similarUnverified在M2 Ultra（最高192GB统一内存）上，MLX可流畅运行700亿参数量级的模型68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/944889API
curl https://kongchang.com/api/v1/knowledge/claims/944889MCP
get_claim(id=944889)