1、 1请参阅附注免责声明DeepSeek R1AIASICDeepseek R1AI1SOC2 Deepseek R1DeepSeek R1o1AI7nmASICDeepseek R1AIAIASIC R1AIDeepseek R1AIAI kV9UiUsVlXqRoM9PbPaQsQpPnPrMkPmMmOfQoPnP9PqQxONZpNmQxNnOtP2请参阅附注免责声明 3请参阅附注免责声明 4请参阅附注免责声明Deepseek R1 API1/tokensDeepseekAIDeepseekAI 5请参阅附注免责声明DeepSeek-V3 Technical ReporttokenFP8

2、 6请参阅附注免责声明DeepSeek-R1:Incentivizing Reasoning Capability in LLMs via Reinforcement LearningR1-zeroo1DeepSeek-R1O1 7请参阅附注免责声明 8请参阅附注免责声明tier 118tier 2tier 1tier 3VEUtier 1AI25%7%tier 3D516nm2521416DRAMAI 9请参阅附注免责声明-20%-10%0%10%20%30%40%050100150200250300350400450YoY7nm5nm3nm2nm晶体管密度（亿晶体管密度（亿个个/平方毫米）

3、平方毫米）0.971.72.5预计2.9（相比3nm增加15%）台积电产能预计台积电产能预计16万片/月20万片/月12.5万片/月25年下半年量产，预计到25年年底月产能达5-6万片，并有望在26年实现产能翻倍257nm 10请参阅附注免责声明 11请参阅附注免责声明DeepSeek AI TCOA100H20H800H100#4444GPU#10,00030,00010,00010,00060,000NVDA$13,50012,50020,00023,00046,000/GPU$23,71624,22831,72834,72879,672$m2377273173471,281$m15738

4、71702307154TCO$m/hr3951,1144875771,996Semi Analysis,DeepSeek558 12请参阅附注免责声明DeepSeekDeepSeekDeepSeek1101000token1000token750DeepSeekToken108*10*1000=1012DeepSeek6710;DeepSeek2*6710*108*1012)/24*60*60=1.6*1019TOPs;FP8H10050%DeepSeekH100(1.6*1019)/(1978*1012*50%)=16,177;DeepSeekA100(1.6*1019)/(624*1012

5、*50%)=51,282;5,CoTDeepSeekH10016177*5=80,885DeepSeekA10051,282*5=256,410370DeepSeekH10016177*370/6710=892DeepSeekA10051282*370/6710=2828DeepSeek110 Token 1,000DeepSeek-R16,710 DeepSeek 1.6*1019TOPsFP850%Deepseek H100 16,177Deepseek A100 51,282CoTDeepseek H100 80,885Deepseek A100 256410370 Deepseek H

6、100 892Deepseek A100 2,828CoTDeepseek H100 4,460Deepseek A100 14,140 13请参阅附注免责声明ASICASICCPUGPUASICASICGPUTPUv5Trainium2H10070%60%,ASIC 14请参阅附注免责声明 15请参阅附注免责声明模型模型参数参数(B)VRAM 需求需求(GB)推荐推荐 GPUDeepSeek-R1-Zero671B1,342 GB多GPU设置（例如，NVIDIA A100 80GB x16）DeepSeek-R1