您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：国产GPU更要拼系统效率与持续交付。
在实测场景中达到H200的性能水平后，国产GPU的竞争，开始进一步转向系统效率、软件适配与实际交付。
本周的天数智芯媒体开放日上，最新一代天垓300与H200的实机对比，是最直观的性能展示。让GPU与专用芯片组成异构系统，是对推理效率的新探索。软件栈引入Agent与Skills，则试图解决客户部署、适配和优化的工程难题。
天数智芯表示，今年完成的一组第三方测试中，在典型PD分离部署下，天垓300折算单卡端到端吞吐达到H200的1.05倍。这表明国产通用GPU在特定高端推理负载中，已经拥有与H200直接竞争的性能基础。
但从性能接近到客户采购，再到持续扩容，中间仍隔着系统效率、软件适配和稳定交付。当硬件具备竞争力，国产GPU怎样把这种能力转化为客户持续使用的理由？当推理催生更多专用芯片，通用GPU的独特价值是什么？
从云端天垓300到端侧彤央系列，天数智芯也在拓展产品布局，将竞争延伸到不同形态的算力需求中。
天垓300对比H200实测，性能小幅领先
过去很长一段时间，峰值算力是国产AI芯片受到关注的重要指标。但对最终用户而言，设备每秒真正产出多少有效Token，比参数表上的理论算力更接近采购决策。
天数智芯披露的第三方测试覆盖四大主流模型、三类场景。其中，1.05倍指典型PD分离部署下，折算单卡的整体端到端吞吐，反映的是特定模型与部署条件下的系统表现。
另外，根据天数智芯性能实验室的数据，在GLM-5.3的64K长上下文场景下，保障同等用户体验时，天垓300的MoE吞吐达到H200的1.15倍。在8K至128K长上下文的Agent场景中，首字延迟较H200降低约10%至30%。
前者关系到系统能够承接多少请求，后者关系到用户发出请求后需要等待多久，它们对应不同工作负载与性能指标。
天垓300能够在这些场景中取得性能领先，一个关键是针对实际负载进行了架构优化。
基于天枢架构的天垓300，单卡显存容量达到144GB，并围绕Attention、MoE等计算环节进行优化。长上下文需要容纳更多状态，也增加了Attention计算压力。MoE虽然通过激活部分专家控制计算量，却带来了小规模矩阵计算、专家分配和数据交换等新问题。
为此，天垓300让Tensor Core与Vector Core并行处理Attention中的矩阵运算和Softmax等向量操作，同时针对MoE中的小规模矩阵、低精度计算与通信协同进行优化。
这些针对性设计，有助于减少不同计算环节之间的等待，把更多硬件资源转化为实际产出。
此次披露的测试与现场演示，展现了天垓300进入高端推理实际竞争区间的能力。对客户而言，下一步要比较的，是在同样的响应要求下，系统能承接多少请求、持续运行表现如何，以及需要付出多少设备、电力和运维成本。
推理需求催生专用芯片，GPU的价值会被削弱吗？
随着推理需求快速增加，以LPU为代表的专用推理芯片备受关注，不同计算环节对硬件的要求也进一步分化。
Prefill处理输入，通常更偏向计算密集型任务；Decode逐步生成Token，更容易受到访存、任务启动和同步开销的影响。进一步拆开，Attention与FFN对计算、带宽和存储的需求也不相同。
当某类任务足够稳定、规模足够大，专用芯片就有机会通过针对性设计，获得更好的性能或能效。让一种通用GPU处理所有环节，未必始终对应最低的成本。
天数智芯展示的AFD分离方案，就是对这种分工的回应。
它在Prefill与Decode分离的基础上，继续拆分Decode阶段的Attention与FFN：天垓300承担Prefill全流程计算，以及Decode中的Attention；合作厂商的XPU承担FFN专家网络计算。
运行过程中，GPU把中间激活数据发送给XPU，XPU完成计算后返回结果，继续推进模型生成。天数智芯给出的对比结果显示，在所展示的方案中，Decode推理延迟降低60%。
天数智芯AI与加速计算负责人单天逸表示：“未来的AI算力系统，不一定由一种芯片包打天下。关键是如何把它们高效组织起来，让整个系统发挥更大的价值。”
异构分工为不同芯片提供了合作空间，GPU的通用性则关系到系统能否持续适应模型变化。
天数智芯副总裁邹翾对雷峰网(公众号：雷峰网)表示，专用硬件可以在固定工作负载中实现较大的性能提升，但模型与算法仍在快速迭代，公司坚持以通用框架为基础，结合专用加速能力。
客户采购设备后，需要持续使用，模型结构、算子和精度需求却可能不断变化。GPGPU平台的通用性提供了继续适配的空间，减少设备与单一计算路径绑定的风险。
强化学习、科学计算和物理仿真，也让计算需求超出了单一推理任务。天数智芯此次展示这些应用，同样体现了GPGPU通用架构更广的使用范围。
推理时代，专用芯片可能分走GPU的部分计算任务，GPU仍能凭借通用性参与更广泛的计算。对国产GPU而言，找到高效分工的位置，通过互联、调度和软件巩固这种价值，成为持续吸引客户的关键。
从性能展示走向商业化，借AI降低软件门槛
异构分工越细，软件需要承担的工作就越多。硬件提供了选择，软件决定这些选择能否变成可交付的服务。
据天数智芯介绍，其已服务380多个客户、覆盖30多个行业。按照积累的项目经验，客户应用迁移通常在小时以内，性能优化则可能需要数天至一周。
模型跟进与迁移效率，也是影响客户复购的重要条件。为此，天数智芯的软件栈引入了Agent与Skills，降低部署和优化过程中的工程成本。
Skills把经过验证的步骤、判断方法和验收要求组织起来，Agent调用已有工具，读取执行反馈，再推进下一步。原本依赖工程师查文档、配环境、读日志和反复试错的工作，有机会部分转化为可复用的流程。
PD分离部署是一个具体例子。它需要分别配置Prefill和Decode实例，传递KV Cache，并打通请求路由与调度。单个组件运行正常，并不意味着整条服务链路已经可用。
按照现场介绍，在所展示的完整模型服务部署案例中，从环境准备到模型正常提供服务，用时由约一小时缩短至约20分钟，人工操作减少为两次目标输入。
进一步，Agent可以结合性能工具提供的轨迹与指标，协助判断瓶颈来自计算、访存、同步还是调度，再选择优化动作。需要修改Kernel时，则通过与FlyDSL协同，生成候选实现，完成编译、正确性检查和目标硬件上的性能测试。
在软件栈中引入AI的重要意义，是让原厂专家的经验能够服务更多开发者。客户规模扩大之后，如果每个项目都依赖专家逐一处理，交付和支持能力就容易成为增长瓶颈。
天数智芯副总裁石加圣强调：“编译器、运行时、算法库和性能工具仍然是技术底座。”
Agent让这些积累更容易调用，也需要依靠它们提供真实反馈。对国产GPU而言，软件研发方式的变化，有机会缩短新模型适配与性能优化的周期，并将这些能力更快交到客户手中。
端侧600T算力的下一代产品已在路上
云端之外，天数智芯此次也展示了彤央端侧产品，并将英伟达Thor T5000作为现场实测对比对象。
天数智芯展示的测试结果显示，在Qwen3.8-27B、Qwen3.6-35B-A3B等大语言模型中，彤央TY1200的性能达到Thor T5000的1.2倍。在英伟达开源的世界基础模型Cosmos-R2-8B测试中，其性能达到Thor T5000的1.8倍。
彤央TY1200提供300 TOPS的INT8稠密算力。与云端产品类似，端侧竞争也在从标称算力，转向具体模型的运行表现，以及设备能否满足现场业务需求。
现场，TY1200展示了离线视频生成、合同审核和物理仿真等应用。其中，在1056×608分辨率下，生成一段5秒视频用时约6分钟，全程在本地完成。围绕MicroDuck机器鸭的演示，则展示了在本地终端上进行物理仿真、策略训练与动作验证的能力。
这些应用对应着不同的部署诉求，企业合同和科研资料需要留在本地，机器人需要在现场完成感知与决策，部分模型任务也开始具备端侧运行的条件。
天数智芯彤央产品市场负责人宋远盈表示：“云端GPU厂商做端侧，不是把芯片做小，而是把能力下沉。”云端积累的模型压缩、量化和软硬件协同能力，以及编译器、算子库和性能工具，成为其拓展端侧市场的基础。
目前，彤央系列通过算力模组与算力终端两种形态，分别面向自主集成和成品部署需求。天数智芯还预告了下一代端侧产品彤央TY2000，稠密算力将达到600T，进一步扩展端侧产品的性能空间。
从云端天垓300，到天数智算软件栈，再到端侧彤央系列，天数智芯已经形成从云到端的产品布局。云端承载大规模训练与推理，端侧贴近数据和业务现场，软件则将模型适配与优化能力延伸到不同产品形态。
天垓300与H200的对比，展现了国产GPU在部分高端负载中的竞争能力。彤央与Thor的对比，则将这种竞争延伸到了端侧。
更完整的产品布局，为天数智芯打开了更多应用入口。能否把性能、软件和交付能力持续转化为客户的实际产出，将决定这套从云到端的布局能够走多远。雷峰网
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: The GPU, which is national production, prefers to spell system efficiency and continuous delivery.
After reaching the H200 performance level in the field, the competition of the domestically produced GPU began to shift further towards system efficiency, software suitability and actual delivery.
The most intuitive display of this week's open day for the smart-core media is that of the latest generation, 300 Skyline versus H200. The formation of GPU and a dedicated chip into an isomer system is a new exploration of reasoning efficiency. The introduction of Agent and Skills in the software warehouse is an attempt to address the engineering challenges of client deployment, fit-out and optimization.
The utensils of days indicate that, in a set of third-party tests completed this year, under a typical PD separation deployment, the twilight of 300 per day is 1.05 times greater than H200. This suggests that the GPU already has a direct competitive performance basis with H200 in a particular high-end reasoning load.
However, from performance to client procurement, to continuous expansion, there is still a gap between system efficiency, software suitability and stable delivery. When hardware is competitive, how does the domestically produced GPU translate this capability into a reason for continued customer use? What is the unique value of Universal GPU when reasoning leads to more dedicated chips?
From the cloud sky 300 to the end side of the congeal series, day cores are also expanding the product layout to extend competition to different forms of computing demand.
It's 300 against H200.
Over a long period of time, peak calculus has been an important indicator of interest for national production of AI chips. For end-users, however, the actual output of the equipment per second was effective token, which was closer to procurement decision-making than the theoretical calculation on the parameter table.
The third-party tests disclosed in the day core cover four main mainstream models and three types of scenarios. Of this, 1.05 times the total end-to-end of the converted card, reflecting the performance of the system under specific model and deployment conditions, was deployed under typical PD separation.
In addition, according to data from the day-to-day mental core performance laboratory, in the 64K long context scene of GLM-5.3, 300 times more MoE per day was swallowed up by H200 when the same user experience was guaranteed. In the Agent scene of 8K to 128K long context, the initial delay was reduced by about 10 to 30 per cent compared to H200.
The former relates to the number of requests that the system is able to receive, and the latter to the length of time a user needs to wait after a request is made, and they respond to different workloads and performance indicators.
One of the keys to achieving performance in these scenarios is to optimize the structure for the actual load.
The skyscraper-based skyscraper 300, single VRAM capacity 144GB, optimized around the Attention, MoE, etc. Long context needs to accommodate more and add to the Attention computational pressure. Although MoE controls calculations by activating some of the experts, it raises new issues such as small-scale matrix computing, expert allocation and data exchange.
To this end, Tianjin 300 allowed Tensor Core to work in parallel with Victor Core on matrix and softmax-like vector operations in Attention, while optimizing the synergy between small-scale matrix and low-precision calculations and communications in MoE.
These targeted designs have helped to reduce the waiting between the different computations and translate more hardware resources into actual outputs.
This disclosure test and the live presentation demonstrated the ability of Tianjin 300 to enter the field of actual competition at the high end of the reasoning. For the client, next step is to compare the number of requests that the system can receive, the performance of ongoing operations and the amount of equipment, electricity and transportation costs that will be required.
Will the value of GPU be diminished by reasoned demand for a dedicated chip?
As the need for reasoning has rapidly increased, the dedicated reasoning chip represented by LPU has received much attention and the requirements for hardware have been further polarized in the different computational chains.
Prefill handles input, usually more computing-intensive tasks; Decode gradually generates Token, which is more vulnerable to access, mission start-up and synchronized costs. Further unbundling, Attention and FNN have different needs for computing, bandwidth and storage.
When certain types of missions are sufficiently stable and large, dedicated chips have the opportunity to achieve better performance or energy efficiency through targeted design. Allowing a generic GPU to process all the links does not always correspond to the lowest costs.
The AFRD separation programme presented by the day-to-day smart core responds to this division of labour.
It continued the splitDecode phase of Attention and FPN: 300 days to take over the entire process of Prefille, and attention in Decode; XPU of the co-worker to take over FPN expert network calculations.
During the operation, GPU sends the intermediate activation data to XPU, returns the result after the XPU calculation and continues to advance the model generation. The comparison given by the day genre shows a 60 per cent reduction in Decode reasoning in the programme presented.
The number of days of ICAI and the one-day time run by the accelerated calculator indicates: “The future ICSS is not necessarily a chip pack. The key is how to organize them efficiently and give the system greater value.”
The isomer division provides a cooperative space for different chips, and the interoperability of GPUs is related to the system ' s continued adaptation to model changes.
The Vice-President of Days and Days indicated to the Thunder Peak Network (WeChat Official Account: Thunder Peak Network) that while dedicated hardware could achieve greater performance in fixed work loads, models and algorithms were still evolving rapidly, and companies insisted on a common framework, combining dedicated accelerator capabilities.
After the purchase of the equipment by the client, the need for continuous use may change the model structure, algorithms and precision requirements. Universality of the GPGPU platform provides space for continued adaptation and reduces the risk of equipment being tied to a single calculation path.
Enhanced learning, scientific computing and physical simulation also allow computing needs to go beyond a single task of reasoning. This display of these applications also reflects the wider use of the GPGPU common architecture.
In the age of reasoning, the dedicated chip may be part of the GPU ' s computational tasks, and the GPU can still be involved in a broader calculation with universality. For the national GPU, finding an efficient division of labour, consolidating this value through interconnection, movement control and software, is key to sustaining client attraction.
From performance to commercialization, use AI to lower the software threshold
The more the isomer division, the more work the software will have to undertake. Hardware provides options, and software determines whether these options can be turned into deliverables.
According to a day-long introduction, it has served more than 380 clients and covered more than 30 industries. Based on accumulated project experience, customer application migration is usually within hours, while performance optimization may take days to a week.
Models follow-up and migration efficiency are also important conditions that affect customer buy-backs. To this end, the day-to-day silo has introduced Agent and Skills to reduce the cost of work in deployment and optimization.
Skills organizes certified steps, judgement methods and acceptance check requirements, and Agent calls on existing tools to read implementation feedback and advance next step. Originally relying on engineers to check documents, set environments, log reading and error-repeated work, there was an opportunity to partially convert to reusable processes.
The separate deployment of PD is a concrete example. It will need to configure the Prefil and Decode examples, transmit KV Cache and access the route and schedule of requests. The functioning of individual components does not mean that the entire service chain is available.
According to on-site presentations, in the full model service deployment cases shown, from environmental preparation to normal model service delivery, time was reduced from approximately one hour to about 20 minutes and manual operations were reduced to two target inputs.
Further, Agent can combine the trajectory and indicators provided by the Performance Tool to assist in determining whether bottlenecks come from computing, accessing, synchronization or scheduling, and then opting for optimisation. When Kernel is required to be modified, the candidate is generated in collaboration with FlyDSL to complete the compilation, correctness check and performance tests on target hardware.
The importance of introducing AI in the software warehouse is to enable the experience of the original plant experts to serve more developers. When the size of the client increases, delivery and support capacity can easily become growth bottlenecks if each project relies on experts on a case-by-case basis.
“The compiler, running time, algorithm and performance tools remain the technological base.”
Agent makes these accumulations more accessible and needs to rely on them for real feedback. For the national GPU, changes in the way software is developed offer the opportunity to shorten the cycle of adaptation and performance optimization of new models and to deliver these capabilities to clients more quickly.
The next generation of 600T calculator on the end is on its way.
Beyond the cloud, the day-to-day brain core this time also showcased the cone side of the product and made Thor T5,000 an on-site comparison.
The results of the tests shown in the day core show that in large language models such as Qwen 3.8-27B, Qwen 3.6-35B-A3B, the performance of TY1200 in Zion amounted to 1.2 times that of Thor T5,000. In the World Basic Model Cosmos-R2-8B test, which is open to British Weeda, its performance was 1.8 times greater than that of Thor T5,000.
TY1200 provides 300 TOPS of INT8-intensive computing. Similar to cloud-end products, end-side competition is moving from nominal computing to the performance of specific models and the ability of equipment to meet on-site operational needs.
At the scene, TY 1200 demonstrated the applications of offline video generation, contract clearance and physical simulation. Of these, under 1056 x 608 Z56Z, a five-second video was generated for approximately six minutes, with the complete local completion. A demonstration around the MicroDuck machine duck demonstrated the ability to perform physical simulation, tactical training and operational validation on local terminals.
These applications correspond to different deployment requirements, business contracts and scientific information need to remain locally, robotics need to perform sensory and decision-making on the ground, and some modelling tasks begin to have end-side conditions.
Song Ying, who is the market manager for the day-to-day Zhiqiang products, states: “The cloud-end GPU manufacturer does the end, not the small chip, but the ability to sink.” Model compression of cloud accumulation, quantization and software and hardware synergies, as well as compilers, algorithms and performance tools, form the basis of their extended side markets.
At present, the Consejo series is oriented towards autonomous integration and product deployment requirements, respectively, in the form of a mathematical model and a calculator terminal. The day cores also forecast the next generation of end-side products, TY2000, with a dense calculus of 600 T, further expanding the performance space for end-side products.
From the cloud sky, 300, to the day-to-day silo, to the end-to-side saloon series, the day-to-day scintillation of products has formed. The cloud carries large-scale training and reasoning, with end-side proximity to data and operational sites, while the software extends model suitability and optimization to different product forms.
A comparison of 300 and H200 shows the competitiveness of domestically produced GPUs in some high-end loads. A comparison between Zongo and Thor extends this competition to the end.
A more complete product layout has opened up more applications for day cores. The continued translation of performance, software and delivery capacity into the actual output of clients will determine how far this cloud-to-end layout can go. Thunder Peak Network
Original Ray Peak article, unauthorized reproduction. For further information, see the information below.
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.