IT之家 9 月 22 日消息,在昨日的 2026 人工智能计算大会(AICC2026)上,浪潮信息宣布推出元脑 SD200 Ultra 超节点 AI 服务器与元脑 HC2000 多元算力机组。
据官方介绍,元脑 SD200 Ultra 基于国产 AI 芯片构建,单机可承载运行 2.8 万亿参数的 Kimi K3 大模型。
浪潮表示,SD200 Ultra 的 Token 生成时延首次降至 5.85 毫秒以内,相当于单使用者速度 170 Tokens/s,达到业界平均水平的 5 倍。该机型兼容 10 万亿参数规模前沿模型单机运行。
IT之家注:Token 生成时延指大模型每输出一个词元所需的平均时间,是衡量推理响应速度的核心指标之一。
SD200 Ultra 使用 3D Hyper Mesh 架构,紧耦合 128 芯本土 AI 芯片,给出 8TB 统一编址显存和 64TB 内存。平台使用原生内存语义通信,通信时延低至 0.69 微秒。
该超节点借助全局统一编址、虚拟影子设备映射、跨域地址翻译与路由,达成 GPU 跨主机域的统一寻址访问,构建百纳秒级低时延内存语义互连域。
浪潮称,借助对称内存工艺,SD200 Ultra 首次在基于本土 AI 芯片的超节点上达成 GPU 显存直连,使 GPU 可直接访问远端 GPU 显存。AllReduce 通信时间降低 3.5 倍。
针对 Kimi K3 推理,SD200 Ultra 使用 Kimi K3 构建超级算子智能体,达成通算融合、Tile 级流水的超级算子,将 KDA、Gated MLA、MoE 中的基础算子融合。算子数量降低 10 倍,推理表现提高 3 倍以上。
浪潮同日发布元脑 HC2000 多元算力机组。该机组使用 DirectCom 2.0 极速架构,基于高密液冷 AI 整机柜,搭配 MCIS 推理软件栈,可动态匹配计算负载。官方称同等投资下 Token 产能提高 10 倍。
浪潮将 Agent 时代 AI 计算分为两个方向:Capability AI Compute 能力型智算,支撑突破智能上限;Capacity AI Compute 容量型智算,达成智能充分供给。
浪潮称,前沿模型参数规模已从 DeepSeek R1 的 6710 亿增长到 Kimi K3 的 2.8 万亿,并走向 10 万亿级;上下文从 128K 扩展到 1M 以上;Agent 从单体走向成百上千个协同。
模型权重、KV Cache(键值缓存,用于存储注意力机制中的键和值,减少重复计算)和并发任务规模同步增长,对显存容量、高速互连和并行扩展效率提出更高要求。
复杂 Agent 任务包含大量“推理 — 工具调用 — 反馈 — 重新规划”循环,每轮时延都会累积。浪潮信息称,时延已不只是感受指标,可能影响任务能否及时完成。
复杂 Agent 任务需连续运行数小时甚至数天,期间经历大量模型调用、工具调用和数据交互。任何计算、通信或软件故障都可能中断任务。
本文来源:https://www.ithome.com/1/005/504.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Home > IT News > Industry
IT Home September 22 news, at the 2026 Artificial Intelligence Computing Conference (AICC2026) held yesterday, Inspur announced the launch of the YuanBrain SD200 Ultra hyper-node AI server and the YuanBrain HC2000 multi-computing power cluster.
According to official introduction, the YuanBrain SD200 Ultra is built on domestic AI chips and can carry and run a 2.8 trillion parameter Kimi K3 large model on a single machine.
Inspur stated that the SD200 Ultra’s token generation latency has been reduced to within 5.85 milliseconds for the first time, equivalent to a single user speed of 170 tokens/s, reaching 5 times the industry average. This model is compatible with a single-machine run of 10 trillion parameter scale frontier models.
IT Home note: Token generation latency refers to the average time required for a large model to output one token, which is one of the core indicators for measuring inference response speed.
The SD200 Ultra uses a 3D Hyper Mesh architecture, tightly coupling 128 domestic AI chips, providing 8TB of unified addressed video memory and 64TB of memory. The platform uses native memory semantic communication, with a communication latency as low as 0.69 microseconds.
This hyper-node achieves GPU cross-host domain unified addressing access through global unified addressing, virtual shadow device mapping, cross-domain address translation, and routing to build a low-latency memory semantic interconnection domain on the order of hundreds of nanoseconds.
Inspur claims that using symmetric memory technology, the SD200 Ultra achieves direct GPU video memory connection for the first time on a hyper-node based on domestic AI chips, allowing GPUs to directly access remote GPU video memory. The AllReduce communication time is reduced by 3.5 times.
For Kimi K3 inference, the SD200 Ultra uses Kimi K3 to build a super-operator agent, achieving compute-communication fusion and tile-level pipeline super-operators, fusing basic operators in KDA, Gated MLA, and MoE. The number of operators is reduced by 10 times, and inference performance is improved by more than 3 times.
Inspur also released the YuanBrain HC2000 multi-computing power cluster on the same day. This cluster utilizes the DirectCom 2.0 high-speed architecture, based on a high-density liquid-cooled AI full rack, paired with the MCIS inference software stack, capable of dynamically matching computing loads. The official claims a 10-fold increase in token productivity under the same investment.
Inspur divides Agent-era AI computing into two directions: Capability AI Compute, which supports breaking through the upper limits of intelligence, and Capacity AI Compute, which achieves a sufficient supply of intelligence.
Inspur states that the parameter scale of frontier models has grown from DeepSeek R1’s 67.1 billion to Kimi K3’s 2.8 trillion, and is heading towards the 10 trillion level; context has expanded from 128K to over 1 million; agents have evolved from single entities to hundreds or thousands of collaborators.
The growth of model weights, KV Cache (key-value cache, used to store keys and values in the attention mechanism to reduce redundant calculations), and the scale of concurrent tasks have raised higher requirements for video memory capacity, high-speed interconnection, and parallel expansion efficiency.
Complex agent tasks involve many “inference — tool invocation — feedback — replanning” loops, where latency accumulates with each round. Inspur claims that latency is no longer just a perceptual metric but can affect whether the task can be completed in a timely manner.
Complex agent tasks require continuous operation for hours or even days, involving a large number of model calls, tool invocations, and data interactions. Any computing, communication, or software failure can interrupt the task.




