IT之家 10 月 28 日消息,天数智芯今日在上海举行 2026 媒体开放日,首次公开展示了天垓 300 通用 GPU 的实机运行。
据天数智芯官方介绍,天垓 300 成为国内首个实测表现达到英伟达 H200 水平的通用 GPU,第三方实测折算单卡总体吞吐达到 H200 的 1.05 倍(IT之家注:具体的测试方法、模型负载和工作负载条件尚未同步公布)。
天垓 300 的实机演示是本次开放日最受关注的部分。在此之前,天数智芯对这款机型的公开信息多集中于架构和路线图层面,实机运行尚属首次。
与天垓 300 配套的超节点思路 IXL144 在稍早的工博会上已经亮相。该思路在单机柜内集成 144 颗 GPU 与 18 颗 CPU,配备约 10.1TB 显存,天数智芯维度称其为“业界最高密度单机柜 GPU 超节点”。据天数智芯测算,该思路每百万有效 Token 的成本较国际主流思路可降低 40%。
本次开放日还披露了异构部署维度的进展:国内首个 GPU 与 XPU 异构部署的 AFD 分离思路成功跑通,推理延迟降低 60%。AFD 分离通常指将注意力、前馈网络等计算阶段在异构算力单元之间开展拆分调度,以匹配不同硬件的擅长负载。
端侧机型线维度,彤央 TY1200 的实测表现被称为“超越 Thor T5000”,并称首次在国产端侧机型上达成了基于 MuJoCo-Warp 的 GPU 加速物理仿真。彤央 TY1200 此前公开的规格为 300 TOPS 稠密算力,面向 AI PC 与具身智能情形。
相关阅读:
本文来源:https://www.ithome.com/1/019/056.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to IT Home reader HH_KK for the tip!
IT Home, October 28 — Tianshu Zhixin today held its 2026 Media Open Day in Shanghai, where it demonstrated for the first time a live, on-hardware run of the Tianhua 300 general-purpose GPU.
According to Tianshu Zhixin, the Tianhua 300 has become China’s first general-purpose GPU whose measured performance reaches NVIDIA H200 levels, with third-party testing indicating that a single card’s overall throughput is equivalent to 1.05 times that of the H200 (IT Home note: the specific testing methodology, model workloads, and workload conditions have not yet been disclosed).
The live demo of the Tianhua 300 was the most closely watched part of the open day. Until now, Tianshu Zhixin’s public information about this model has mostly focused on architecture and roadmap-level details, with an actual on-hardware run being the first of its kind.
The IXL144 supernode concept paired with the Tianhua 300 had already been unveiled earlier at the China International Industry Fair. The concept integrates 144 GPUs and 18 CPUs within a single rack, with roughly 10.1TB of total graphics memory capacity; Tianshu Zhixin said it is “the industry’s highest-density single-rack GPU supernode.” Tianshu Zhixin estimates that the concept can reduce the cost per one million valid Tokens by 40% compared with mainstream international approaches.
The open day also disclosed progress in heterogeneous deployment: China’s first heterogeneous GPU + XPU AFD disaggregation concept has been successfully implemented, reducing inference latency by 60%. AFD disaggregation generally refers to splitting and scheduling compute stages such as attention and feed-forward networks across heterogeneous computing units so that each workload runs on the hardware best suited to it.
On the edge-device product line side, the measured performance of the Tongyang TY1200 is claimed to “surpass Thor T5000,” and the company said it has achieved, for the first time on a domestic edge model, GPU-accelerated physics simulation based on MuJoCo-Warp. Previously disclosed specifications for the Tongyang TY1200 include 300 TOPS dense computing performance, targeting AI PC and embodied-intelligence scenarios.
Related reading:

