Taichu Yuanqi’s Hypertintellix Named “Computing China Annual Outstanding Achievement” at 2026 China Computing Conference

The three-day conference is being held in Langfang, Hebei, and runs alongside the “Computing China · AI and Computing Industry Exhibition,” which gathers domestic companies for first-time technology and product launches. Taichu Yuanqi was one of the companies making a debut, showing a distributed computing solution and, for the first time offline, a 20-foot standardized containerized computing product.

Hong Yuan, chief product officer and senior vice president of Taichu Yuanqi, said computing demand is growing exponentially as computing networks expand and AI is applied across industries, while the domestic computing gap is widening. To meet demand from different industries more flexibly and support multi-point computing network construction, the company developed a distributed computing solution using 20-foot and 40-foot standardized containers as carriers.

Hong said the containerized computing units can be configured with 32 to 256 domestic AI accelerator cards, use air-cooled or liquid-cooled designs, and support horizontal expansion into thousand-card computing clusters. Different combinations can deliver up to 80 PFLOPS of FP16 computing output in a single container, he said, meeting demand for large-model training and inference as well as scientific simulation.

The 20-foot unit displayed for the first time carries Taichu Yuanqi’s self-developed TC-64-1108 air-cooled intelligent computing server rack, with 64 domestically developed AI accelerator cards and one management network cabinet. It uses inter-column air conditioning plus air cooling and reaches 20 PFLOPS of FP16 computing power.

The company said computing demand is spreading from central clouds to edge environments, with small- and medium-scale demand rising in sectors such as finance and energy. Building new intelligent computing centers takes time and high investment, making it difficult to meet vertical-industry computing needs quickly. The distributed solution combines prefabrication with computing: rack assembly, liquid-cooling pipe debugging and equipment testing are standardized in the factory, with a prefabrication rate above 90 percent.

After leaving the factory, the unit needs only a power connection on site and can be deployed within 24 hours, with installation completed in as little as two hours, the company said. That is 70 percent more efficient than traditional cluster delivery, according to the report. The modular, building-block design allows on-demand configuration and raises resource utilization by more than 60 percent compared with phased expansion of traditional intelligent computing centers.

The containerized computing units can also be placed near solar, wind and other renewable power stations to consume green electricity locally and coordinate computing with power. Similar in form to emergency communications vehicles, they are intended for flexible scheduling and can be used by research institutions, industrial bases, emergency computing support and regional computing nodes, the company said. Hong said Taichu Yuanqi had worked with ecosystem partners such as Hanteng Technology on technical co-creation and deployment validation to refine the product for rapid application.

In addition to the distributed solution, Taichu Yuanqi brought a full-stack computing portfolio to the conference, including its new-generation domestic AI heterogeneous chip T2Ultra, the super-intelligent integrated computing system, the domestic AI4S computing platform and the desktop AI product Yuanqi Mini Station for enterprises and developers.

The Hypertintellix system named for the award is suited to both trillion-parameter large-model training and inference and AI4S scientific computing. The unit displayed at the event is a 256-card cluster that uses a high-speed optical interconnect architecture and can be smoothly expanded to a 100,000-card computing scale, according to the company. It supports accelerated and autonomous computing modes, natively supports FP4 to FP64 full-precision computing, and relies on deep software-hardware coordination to reduce host resources and lower research and AI engineering costs, the company said. It is intended to support high-end HPC research tasks and AI applications including large models, multimodality and agents, forming a fully self-controllable high-end integrated computing base.

The AI4S computing platform is a one-stop research innovation platform developed by Taichu Yuanqi for fields such as weather forecasting, biomedicine, quantum mechanics, chemical materials and fluid mechanics. It uses self-developed heterogeneous many-core AI chips as its computing base and builds a full-stack native research solution covering hardware, basic components, acceleration libraries, model frameworks and development kits, the company said. The platform is compatible with mainstream CPUs including Loongson, Sunway, Phytium and x86, and provides a self-developed programming model and acceleration libraries for general operators and scientific computing. It also adapts to PyTorch, JAX, PaddlePaddle and MindSpore, and includes a DevKit for operator generation, compilation optimization and performance analysis, lowering the threshold for operator development and tuning.

Taichu Yuanqi said it has worked with Tsinghua University, Hunan University, Shandong University, Baidu, Dongrun and other university and industry partners on efficient and accurate research solutions in weather forecasting, quantum simulation, gene analysis, material simulation and fluid computing. The article was provided by Taichu Yuanqi and reposted by QbitAI with authorization, with views belonging to the original author.