[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-biren-npo-1024-supernode":3,"news-related-5b5cd125-c167-4472-8dc5-2414699b73ce":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"5b5cd125-c167-4472-8dc5-2414699b73ce","壁仞科技 NPO 1024卡超节点:把国产 GPU 集群从'集中'拆成'解耦'","WAIC 2026 上,壁仞科技正式推出下一代 NPO 光互连、分布式解耦架构超节点方案,把单个超节点的 Scale-up 容量拉到 1024 卡。这是继 2025 年 WAIC 联合曦智、上海仪电、中兴发布'光跃 LightSphere X'并拿下 SAIL 奖之后,壁仞在光互连赛道的第二次代际跃迁。\n\n这次迭代的关键字是'分布式解耦'。在传统超节点方案里,GPU 数量受限于单机柜功耗天花板,要么堆密度,要么堆机柜。LightSphere X 已经用硅光 + 分布式光交换 (dOCS) 把单机柜与算力解耦,2026 的 NPO 方案则更进一步:把'集中式 OCS'拆成嵌入每张 GPU 的光交换模块,拓扑可实时重构,故障秒级切换,模型按需动态裁剪超节点规模。\n\n直接收益有三层:第一,Scale-up 容量提升后,留给 Scale-out 的带宽压力随之下降;第二,光交换协议不再绑定特定厂商,绕开了 NVLink\u002FInfiniBand 的生态锁定;第三,硅光芯片不依赖最先进工艺,供应链可控性更强。\n\n把视角拉到中国 AI 算力大盘:从 LightSphere X 的 2 千卡到 NPO 的 1024 卡 Scale-up,壁仞、曦智、中兴、仪电这条国产光互连链已经把'集群'与'超节点'的边界重新画过一遍。当上层模型动辄进入万亿参数、多模态长上下文时代,这种'协议开放 + 硬件解耦 + 软件自主'的栈,是少数能让国产 GPU 不被网络层卡脖子的路径之一。","https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3900457432680067","c51aeeb3-f0b8-4e97-ab9e-da2be69c5f8d",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":21,"name":22,"slug":22,"description":13,"color":13},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"91ea44ed-99d5-4764-9367-16a3d9551069","en","Biren's NPO 1024-GPU supernode decouples the cluster","At WAIC 2026, Biren Technology officially launched its next-generation NPO optical interconnect, distributed-decoupled-architecture supernode solution, pulling the single-supernode Scale-up capacity to 1024 cards. This is the second generational leap for Biren in the optical interconnect track after partnering with Xizhi, Shanghai YiDian, and ZTE at WAIC 2025 to release \"LightSphere X\" and take the SAIL award. This iteration's key word is \"distributed decoupling\". In traditional supernode solutions, the GPU count is limited by the power ceiling of a single rack — you either stack density or stack racks. LightSphere X has already decoupled the single rack from compute using silicon photonics + distributed optical circuit switching (dOCS); the 2026 NPO solution goes one step further: it splits the \"centralized OCS\" into optical switching modules embedded in each GPU, with topologies that can be reconfigured in real time, fault-level failover in seconds, and on-demand dynamic trimming of supernode scale per the model. Direct benefits across three layers: first, the Scale-up capacity boost lets the bandwidth pressure left for Scale-out fall accordingly; second, optical switching protocols are no longer bound to specific vendors, sidestepping the NVLink\u002FInfiniBand ecosystem lock-in; third, silicon photonics chips don't depend on the most advanced processes, giving a much stronger supply-chain controllability. Zooming out to China's AI compute landscape: from LightSphere X's 2,000 cards to NPO's 1024-card Scale-up, the Biren \u002F Xizhi \u002F ZTE \u002F YiDian domestic optical interconnect chain has redrawn the boundary between \"cluster\" and \"supernode\". As upper-layer models routinely enter the trillion-parameter and multimodal-long-context era, this stack of \"open protocol + hardware decoupling + software autonomy\" is one of the few paths that can keep domestic GPUs from being strangled at the network layer.","biren-npo-1024-supernode","2026-07-19T02:30:00Z","2026-07-19T02:05:01.600555Z","2026-08-19T02:08:40.142862Z",true,"agent",143,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"fb1cbe25-8b85-41ec-b619-9a27b405ec34","AMD 收购 Taalas:把 AI 模型权重「刻进硅片」的推理新打法","amd-acquires-taalas-inference-chip","2026-08-19T01:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"5f5bd5f2-9a02-470b-aa25-3f27fb9bb093","字节跳动正训练 10 万亿参数模型，规模对标 Anthropic Mythos 5","bytedance-10t-parameter-model-ft","2026-08-07T09:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"b2c169c6-5150-4423-8073-bf480a2d8745","腾讯 UniPert-G2CP 登《Cell》主刊：把基因扰动和化学扰动塞进同一个语义空间","tencent-unipert-g2cp-cell-virtual-cell","2026-07-31T07:49:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"ae924988-53ef-4f46-b1c2-c43c4e9866fe","欧盟掏 100 亿欧元建 7 座 AI 超级工厂：每座堆 10 万颗顶尖芯片，瞄准美国算力代差","eu-ai-gigafactory-7-factories-100000-chips","2026-07-31T04:30:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"bb1d01e1-c61f-4428-b91a-41000a26bec5","从「堆硬件」到「卖Token」:10余家上市公司押注Token工厂,算力行业 TaaS 模式浮出水面","china-ai-token-factory-taas-shift-2026","2026-07-31T04:00:00+00:00"]