{"id":81150,"topic":"ai","source":"tech.sina.cn","title":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU - tech.sina.cn","url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","url_hash":"6888733d6175b39a98cfc1e1f543a5c20139028e","author":"","summary":"<a href=\"https://news.google.com/rss/articles/CBMilAFBVV95cUxNd00zRU41YjNRTS03cVNZUXU3YTcxaWhnTWZ2Yjhhc2NMQWVjR3VvQWMxWkRLSk5RVW43OFU0ZTA5SWNMZ3g4VGk1aG1jR000VUg5cUl2RGUwSzZUd1FGaG85UERodDNnTUpnZzRLVEQ0a0lsRkhWMTV4NEZUTmNVdl81SERCVFpmZjN0cmlZb3czUzcx?oc=5\" target=\"_blank\">中国移动云发布异构混合推理系统：类脑芯片+国产GPU</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">tech.sina.cn</font>","content":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","image_url":"https://n.sinaimg.cn/spider20260913/200/w600h400/20260913/761d-2e296e2cc2fa2111ca92c71d1d1c08df.png","lang":"zh","published_at":"2026-09-13T08:04:08+00:00","fetched_at":"2026-09-13T08:15:08+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","cluster_id":null,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 642 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":642,"summary_length":642,"usable_text_length":642,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":642,"summary_length":642}},"news_item":{"id":81150,"canonical_url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","source_url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","title":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU - tech.sina.cn","source_name":"tech.sina.cn","author":null,"published_at":"2026-09-13T08:04:08+00:00","locale":"zh","topic":"ai","tags":[],"rss_summary":"<a href=\"https://news.google.com/rss/articles/CBMilAFBVV95cUxNd00zRU41YjNRTS03cVNZUXU3YTcxaWhnTWZ2Yjhhc2NMQWVjR3VvQWMxWkRLSk5RVW43OFU0ZTA5SWNMZ3g4VGk1aG1jR000VUg5cUl2RGUwSzZUd1FGaG85UERodDNnTUpnZzRLVEQ0a0lsRkhWMTV4NEZUTmNVdl81SERCVFpmZjN0cmlZb3czUzcx?oc=5\" target=\"_blank\">中国移动云发布异构混合推理系统：类脑芯片+国产GPU</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">tech.sina.cn</font>","full_text":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","excerpt":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 642 characters.","diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html%3Ffromtech%3D1%26vt%3D4%26cid%3D79649%26node_id%3D79649","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 642 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":642,"summary_length":642,"usable_text_length":642,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":642,"summary_length":642}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU - tech.sina.cn","url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","summary":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","source":"tech.sina.cn","date":"2026-09-13T08:04:08+00:00","content":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html%3Ffromtech%3D1%26vt%3D4%26cid%3D79649%26node_id%3D79649","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 642 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 642 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":642,"summary_length":642,"usable_text_length":642,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":642,"summary_length":642}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/81150","export_markdown":"/api/items/81150/export?format=markdown","export_json":"/api/items/81150/export?format=json","diagnose":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html%3Ffromtech%3D1%26vt%3D4%26cid%3D79649%26node_id%3D79649"},"formats":{"full":{"id":81150,"title":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU - tech.sina.cn","url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","source":"tech.sina.cn","author":null,"published_at":"2026-09-13T08:04:08+00:00","locale":"zh","topic":"ai","tags":[],"excerpt":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","full_text":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","reading_time_min":1,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 642 characters.","diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html%3Ffromtech%3D1%26vt%3D4%26cid%3D79649%26node_id%3D79649","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 642 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":642,"summary_length":642,"usable_text_length":642,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":642,"summary_length":642}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 642 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":642,"summary_length":642,"usable_text_length":642,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":642,"summary_length":642}},"actions":{"read":"/item/81150","export_markdown":"/api/items/81150/export?format=markdown","export_json":"/api/items/81150/export?format=json","diagnose":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html%3Ffromtech%3D1%26vt%3D4%26cid%3D79649%26node_id%3D79649"}},"digest":{"id":81150,"title":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU - tech.sina.cn","url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","source":"tech.sina.cn","topic":"ai","published_at":"2026-09-13T08:04:08+00:00","excerpt":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU 快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。 该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。…","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 642 characters.","reading_time_min":1,"cluster_id":null},"card":{"display_title":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU - tech.sina.cn","subtitle":"tech.sina.cn · 2026-09-13","summary":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU 快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。 该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。…","badges":["quality:high"],"links":{"read":"/item/81150","original":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","diagnose":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html%3Ffromtech%3D1%26vt%3D4%26cid%3D79649%26node_id%3D79649"},"quality_warning":null},"export":{"title":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU - tech.sina.cn","url":"https://finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html?fromtech=1&vt=4&cid=79649&node_id=79649","summary":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","source":"tech.sina.cn","date":"2026-09-13T08:04:08+00:00","content":"中国移动云发布异构混合推理系统：类脑芯片+国产GPU\n快科技9月13日消息，2026中国算力大会在河北廊坊举办。在大会“算力首创首发”专场上，国内首个国产GPU＋类脑芯片大模型异构混合推理系统正式发布。\n该系统由移动云（中移苏州软件技术有限公司）联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同研发，目标是让国产芯片高效支撑大模型推理。\n其技术思路是将大模型拆解分工。团队对Transformer结构做PD/AF分离：Prefill预处理阶段与Attention注意力计算交给国产GPU，FFN模块交给类脑芯片。类脑芯片的存算一体和片上大容量SRAM架构，正好承接对带宽敏感的部分。\n研发团队同时自研了模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度与结果聚合。这套方案对标英伟达下一代Vera Rubin with Groq异构推理架构。\n实测数据落在硬件配置上：3台天数智芯GPU服务器搭配3台类脑机柜，运行DeepSeek V4 Flash模型，相较同等投入规模的纯GPU集群，推理输出和推理能效均提升1倍以上，业务运营成本下降40%以上。\n项目累计形成15项授权发明专利、6项软件著作权，覆盖异构调度、类脑芯片架构、编译工具链等环节，目前已进入小批量试产阶段。\n该方案面向Token工厂、AI代码生成、多智能体协同等场景，也可用于金融、安防、通信等安全敏感行业。\n【本文结束】如需转载请务必注明出处：快科技\n责任编辑：鹿角","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-13/detail-inirskfx0431159.d.html%3Ffromtech%3D1%26vt%3D4%26cid%3D79649%26node_id%3D79649","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 642 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 642 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":642,"summary_length":642,"usable_text_length":642,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":642,"summary_length":642}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}