Tecorigin T2 Ultra and HyperIntelliX: A Unified AI and HPC Computing Platform

Introduction

WAIC 2026 turned Shanghai into a showcase for China’s rapidly expanding AI-computing industry. The exhibition brought together 108 chips and more than 200 intelligent-computing products, covering accelerators, supernodes, cluster systems, software stacks, scientific-computing platforms, and industry solutions.

The tone of the competition has changed.

In earlier years, vendors often led with the peak performance of a single card. This year, the recurring themes were system architecture, large-scale coordination, interconnect efficiency, software migration, and complete deployment capability.

As model sizes grow, agents become more stateful, and AI for Science workloads demand both general-purpose and accelerated computing, the value of an isolated chip becomes harder to separate from the system around it.

At Tecorigin’s booth in Zhangjiang, the company presented a new heterogeneous many-core chip and a large supernode system designed around what it calls AI-and-HPC convergence.

The core products included:

• The HCU heterogeneous converged computing architecture
• The new T2 Ultra processor
• The HyperIntelliX large-scale computing system
• An upgraded developer community and migration toolchain
• A domestically built AI4Science computing platform
• A multi-party, full-domain large-model security system

The company’s argument is straightforward: the market is no longer deciding only whether a domestic chip can run a model. It is deciding whether the complete platform can support different workloads, scale to large clusters, reduce migration costs, and remain stable in production.

Architecture Comes Before Single-Card Performance

Tecorigin CEO Yang Jinzhe described the industry as moving from single-card competition to system-level competition.

That shift reflects changes in the workloads themselves.

Large models have crossed the trillion-parameter threshold. Agent systems require more planning, scheduling, memory management, and tool coordination. AI4Science applications add high-throughput numerical computing, long-running jobs, and strict double-precision requirements.

Simply adding more accelerator cards can produce diminishing returns when the surrounding CPUs, memory, communication links, storage, and scheduling software cannot keep up.

The relevant objective is no longer just maximum theoretical compute. It is reducing the cost of useful compute across the complete system.

Other WAIC exhibitors presented similar system-level strategies, including Huawei’s Atlas 950 SuperPoD, MetaX’s Xijing supernode systems, and Alibaba’s Lingjun Zhenwu infrastructure.

Tecorigin’s answer is its HCU architecture, short for heterogeneous converged computing.

The architecture places general-purpose CPU resources and dedicated AI-computing cores on the same high-speed bus. It supports a reconfigurable M:N ratio, allowing the balance between CPU and accelerator resources to change with the workload.

In the proposed division of labor:

• The CPU side handles agent scheduling, workflow control, data preprocessing, and other general-purpose tasks.
• The XPA computing array concentrates on large-model training and inference.
• Shared hierarchical memory manages hot and cold data dynamically.
• The system changes the resource balance according to the task rather than relying on one fixed CPU-to-accelerator configuration.

In practical terms, the design combines two types of computing resources inside one architecture and allows the system to allocate them more flexibly.

The potential advantage is better utilization across mixed workloads. A system does not need to use a high-throughput AI array for every control-heavy task, nor depend on a relatively small host CPU when the workload requires extensive scheduling and coordination.

Why Heterogeneous Many-Core Computing Matters for Agents

Traditional AI servers commonly use one CPU with four to eight GPUs or AI accelerator cards.

In that design, the CPU often serves as a host. It launches tasks, manages data movement, handles operating-system work, and coordinates the accelerators.

This model works well when the accelerator performs most of the computation and the control path remains relatively simple.

Agentic AI changes that balance.

A capable agent may need to:

1. Interpret an objective.
2. Divide it into subtasks.
3. Select tools.
4. Plan and revise a workflow.
5. Read and write external data.
6. Maintain long-term memory.
7. Coordinate several models or subagents.
8. Recover from failed actions.
9. Manage permissions and approvals.
10. Continue execution over a long session.

Many of those operations depend on general-purpose computing rather than dense matrix multiplication.

The source article cites an industry expectation that the historical one-to-many CPU-to-GPU ratio may move closer to one-to-one in some future nodes, with CPU demand potentially rising further for agent-heavy workloads.

Alibaba Cloud has also described agents as combining stateful scheduling with stateless execution, creating a need for more precise resource control.

Tecorigin’s HCU architecture is intended to respond to that shift at the processor level.

By connecting CPU resources and AI cores through the same high-speed fabric, the company aims to reduce data-transfer overhead compared with a fully separated host-plus-accelerator design.

Whether the theoretical advantage becomes a practical one depends on real application performance, compiler quality, memory behavior, scheduling, power efficiency, and system utilization.

T2 Ultra: One Processor, Two Operating Modes

Tecorigin introduced the T2 Ultra as the first processor built around its new HCU architecture.

The most distinctive feature is its dual-mode design.

Autonomous Computing Mode

In autonomous mode, T2 Ultra can operate as a complete computing node rather than functioning only as an attached accelerator.

This makes it relevant to:

• Edge deployments
• Smaller research systems
• Workstations
• Independent compute nodes
• Environments where a separate host architecture would add complexity

Accelerated Computing Mode

In acceleration mode, T2 Ultra works alongside a host system and acts as the high-throughput computing engine inside a larger server or cluster.

This supports:

• Large-model training
• Large-scale inference
• Multi-card servers
• Supernodes
• Scientific-computing clusters
• Data-center deployments

The same processor family can therefore be used in more than one infrastructure role.

Reported T2 Ultra Specifications

Tecorigin and the original report list the following peak specifications:

These are vendor-supplied peak figures.

FP64 performance is especially relevant to scientific and engineering applications that require double-precision arithmetic. FP16 and FP4 target AI training and inference, with lower-precision formats improving throughput when the model and accuracy requirements permit their use.

Sparse-compute figures also assume that the workload and software can exploit the supported sparsity pattern. They should not be compared directly with dense performance without confirming the benchmark method.

The practical value of T2 Ultra will depend on:

• Sustained application performance
• Memory bandwidth and capacity
• Cluster scaling efficiency
• Accuracy at lower precision
• Software support
• Reliability
• Power consumption
• Production availability
• Cost per completed workload

The dual-mode architecture may reduce hardware fragmentation for customers that need several deployment sizes. It does not remove the need to validate each target workload.

HyperIntelliX: Combining Intelligent Computing and Supercomputing

If T2 Ultra is the computing engine, HyperIntelliX is the large-scale platform built around it.

Tecorigin positions HyperIntelliX as a domestic computing system that combines:

• Intelligent computing for model training and inference
• High-performance computing for numerical simulation
• AI4Science workflows
• Large-scale interconnection
• Liquid-cooled modular infrastructure
• Shared software and resource management

The company says the system uses a liquid-cooled modular design, targets a PUE below 1.1, and can expand through optical interconnection toward clusters containing up to 100,000 cards.

It is intended to run both trillion-parameter model workloads and scientific-computing applications on one hardware family.

This is the main meaning of supercomputing and intelligent-computing convergence.

A conventional AI cluster is often optimized for tensor operations, model training, or inference. A conventional supercomputer may prioritize FP64, MPI-style workloads, large numerical simulations, and sustained parallel performance.

AI4Science increasingly combines the two.

A research workflow may use conventional simulation to generate data, train a neural surrogate model, run AI-assisted prediction, and return to numerical validation. Maintaining separate infrastructures for every stage can create duplicated hardware, software, and operations costs.

HyperIntelliX is designed to provide one underlying platform for both categories.

The claim that it is the industry’s first large-scale converged AI-and-HPC platform comes from the company and original report. The classification depends on how competing systems and previous deployments are defined, so it should not be treated as an independently established market ranking.

The Platform Is Designed for More Than Large Models

Tecorigin emphasized that HyperIntelliX is not limited to general AI training and inference.

The company presented AI4Science projects in three representative areas:

• Global weather-prediction visualization
• The TecoQSim classical quantum simulator
• Large-scale virtual drug screening

These workloads impose different demands from a typical chatbot or retrieval application.

Scientific computing often requires:

• High concurrency
• High data throughput
• Long-running jobs
• Stable execution over many hours or days
• Double-precision arithmetic
• Large distributed-memory systems
• Specialized numerical libraries
• Reproducible results
• Integration with established research codes

The Tecorigin team’s background includes participation in systems and projects that received the Gordon Bell Prize, one of the best-known awards in high-performance computing.

That experience supports the company’s decision to target AI4Science, although historical team achievements do not independently validate the performance of a new product.

Life Sciences

Tecorigin’s official AI4Science materials describe work including:

• A domestic reproduction of AlphaFold3 with Baidu PaddlePaddle and the Shenwei team
• Adaptation and optimization of UCSF DOCK 3.7 with Shandong University
• Development of CrossDNA with Hunan University

These projects target protein-structure prediction, molecular docking, virtual drug screening, and genomic-language modeling.

Weather and Climate

Tecorigin and Baidu’s scientific-computing team have worked on a domestic weather system that visualizes temperature, wind, rainfall, and typhoon trajectories.

Weather prediction requires large spatiotemporal datasets and can combine numerical models with machine-learning systems.

Quantum Simulation

Tecorigin and a Tsinghua University team developed TecoQSim, a classical simulator for quantum systems.

The platform uses complex tensor contraction and high-performance numerical computation to simulate quantum behavior on conventional computing hardware.

Fluid Dynamics and Materials

The company also describes work in computational fluid dynamics, scientific multimodal models, materials, chemistry, and other fields that combine simulation and AI.

These partnerships show the intended range of the platform. Performance claims for individual projects remain specific to their configurations and evaluation methods.

Ecosystem Migration Is the Hardest Part of Domestic Computing

Strong hardware is not sufficient if developers cannot move their workloads onto it.

Tecorigin used WAIC to release two ecosystem products intended to reduce migration costs.

The first was AI Developer Community 2.0.

The company says the platform supports major training and inference systems including:

• Megatron-LM
• DeepSpeed
• PaddlePaddle and PaddleHelix
• PyTorch
• vLLM
• SGLang
• Transformers
• LLaMA-Factory

The toolchain is intended to cover:

1. Model analysis
2. Migration
3. Operator development
4. Compilation
5. Performance optimization
6. Accuracy debugging
7. Deployment
8. Runtime management

The second product was a new domestic AI4Science computing platform.

It targets the characteristics of scientific workloads and combines:

• Computing-resource scheduling
• Job management
• Development environments
• Scientific-computing suites
• Tecorigin’s programming models
• General and domain-specific acceleration libraries
• Operator-generation tools
• Compiler optimization
• Performance analysis

According to the official product page, the platform supports mainstream and domestic CPUs, including Loongson, Shenwei, Phytium, x86, and other systems.

It also supports frameworks including:

• PyTorch
• JAX
• PaddlePaddle
• MindSpore
• vLLM

The accompanying DevKit is designed to lower the cost of writing and optimizing operators.

Why Migration Takes Time

A model that starts successfully is not necessarily production-ready.

Migration can require work at several layers:

• Replacing unsupported operators
• Porting custom kernels
• Updating build systems
• Converting model formats
• Verifying numerical accuracy
• Reproducing baseline results
• Improving memory efficiency
• Optimizing communication
• Tuning parallelism
• Testing failure recovery
• Updating monitoring and deployment tools

The source article cites an industry estimate of three to six months for a complete adaptation, with performance optimization accounting for approximately 60% of the work.

That estimate is not universal. A standard model using supported operators may migrate much faster, while a large custom research code can take longer.

The important point is that migration cost is measured in engineering time and operational risk, not just software installation.

Two Broad Ecosystem Strategies

The source article divides domestic accelerator ecosystems into two broad approaches.

Compatibility-Oriented Ecosystems

This route tries to preserve familiar development patterns and minimize changes to existing workloads.

A vendor may provide its own compiler, runtime, operator libraries, and migration tools while supporting widely used frameworks and programming models.

The advantage is lower switching friction.

The risk is that complete compatibility with a mature, fast-changing ecosystem is difficult to maintain. Small semantic or performance differences can appear in advanced workloads.

Independent Ecosystems

The second route builds a more independent software environment, such as the model associated with Huawei’s MindSpore and Ascend stack.

This can provide deeper control and optimization around the vendor’s own hardware.

The trade-off is that developers may need to learn new tools or modify workflows more extensively.

Tecorigin’s public materials place it closer to the compatibility-oriented approach. Its website describes support for mainstream models and frameworks and says its software stack can adapt more than 40 model families.

The source article estimates that the approach can address 80% to 90% of mainstream scenarios. That range is a broad vendor or industry characterization, not a guarantee that every workload will migrate without changes.

Building the AI4Science Ecosystem

Tecorigin’s ecosystem work includes collaborations with universities, research institutes, and companies.

The original article lists examples such as:

• TecoQSim with Tsinghua University
• A quantum-supercomputing-intelligent-computing platform with SpinQ
• RiverONE adaptation and Loongson-platform validation
• Infrastructure support for Tsinghua AIR’s AtomWorld
• AlphaFold3 reproduction with PaddlePaddle and Shenwei teams
• CrossDNA with Hunan University
• Molecular-docking optimization with Shandong University
• AI4Science solutions with Baidu PaddlePaddle
• A domestic weather-computing appliance with Baidu’s scientific-computing team
• Energy-computing work with Dongrun
• Adaptation of Shanghai AI Laboratory’s Intern-S1 scientific multimodal model

Tecorigin’s official website confirms several of these projects and describes more than 40 adapted model families across language, multimodal, code, and AI4Science use cases.

The source also says the company has worked with more than 100 universities, research institutions, and industry partners.

Partnership counts and adaptation totals are company-reported. They show ecosystem activity, but they do not by themselves establish the depth, production scale, or commercial success of each deployment.

The central strategy remains clear:

The chip provides access to the market, but the software and application ecosystem determines whether the platform remains useful.

A Full-Domain Large-Model Security System

WAIC also included the launch of a full-domain large-model security system built by seven organizations:

• Henan Airport Intelligent Computing Center
• RealAI
• Tecorigin
• Anshi Information
• Qingyuan Zhiqian
• Dianshu Technology
• Shuanxin

The source describes the system as a layered architecture.

At the bottom is a domestic trusted-computing foundation. Above it is an integrated evaluation and governance platform. The upper layers cover:

• Large-model security evaluation
• Data compliance
• Data security
• Software supply-chain security
• Security training and education

This may appear separate from processor and cluster performance, but security is increasingly part of infrastructure readiness.

Government, finance, healthcare, and other regulated sectors cannot deploy a model solely because it runs quickly. They need controls covering data, permissions, software components, model behavior, auditing, and compliance.

The value of the combined security system will depend on its standards, coverage, testing methods, integration quality, and independent validation.

A layered diagram is not the same as an operational security guarantee.

A Three-Layer Infrastructure Strategy

The source article summarizes Tecorigin’s WAIC portfolio as three connected layers.

Bottom Layer: Hardware and Computing Systems

• HCU heterogeneous converged computing architecture
• T2 Ultra processor
• HyperIntelliX computing system
• Large-scale interconnection
• Liquid-cooled cluster infrastructure

Middle Layer: Developer and AI4Science Platforms

• AI Developer Community 2.0
• Framework adaptation
• Operator development
• Compiler and performance tools
• AI4Science platform
• DevKit and acceleration libraries

Upper Layer: Security and Industry Applications

• Full-domain large-model security
• Research partnerships
• Weather, quantum, life-science, and materials workloads
• Enterprise and public-sector deployments

The layers are intended to reinforce one another.

A chip without software has limited adoption. Software without reliable hardware cannot deliver performance. Both are difficult to deploy in sensitive industries without governance and security.

Tecorigin is therefore positioning itself not as a supplier of one replacement accelerator, but as a builder of a broader AI infrastructure stack.

WAIC 2026 Was a Test of Domestic Computing Maturity

The source places Tecorigin within a larger domestic-chip market that includes Huawei Ascend, Cambricon, MetaX, Hygon, and other vendors.

Several figures in that comparison—including production targets, revenue growth, market capitalization, and model-adaptation counts—are time-sensitive and come from company statements or financial reports. They are not necessary to evaluate the Tecorigin architecture and may change quickly.

The more durable observation is that domestic AI computing is moving from policy-led adoption toward stronger market testing.

Customers increasingly ask:

• Can the hardware be delivered at scale?
• Can a large cluster remain stable?
• Does the software support current models?
• How long does migration take?
• What is the effective cost per workload?
• Can engineers diagnose problems?
• Does performance hold after model upgrades?
• Is the platform secure enough for regulated data?
• Will the vendor support the system over its full lifecycle?

Tecorigin’s differentiation is its attempt to support AI and scientific computing through one architecture.

The opportunity is clear. AI4Science workloads require both high-precision numerical performance and modern AI acceleration.

The challenge is equally clear. The company must demonstrate:

• Reliable large-scale production
• Stable 10,000-card and 100,000-card deployments
• High utilization
• Strong software compatibility
• Repeatable model adaptation
• Competitive total cost of ownership
• Sustainable technical support
• Independent performance verification

Moving from “usable” to “easy to use” is the decisive step for any alternative computing ecosystem.

The Real Competition Is System Completeness

WAIC 2026 showed that the era of relying on one peak-compute number is ending.

Modern AI infrastructure has to coordinate:

• General-purpose computing
• Accelerated computing
• Memory
• Interconnects
• Storage
• Cooling
• Compilers
• Frameworks
• Operator libraries
• Cluster scheduling
• Model adaptation
• Scientific software
• Security and governance

T2 Ultra is Tecorigin’s new chip-level answer. HyperIntelliX is its system-level answer. The developer platform and AI4Science stack are its ecosystem answer. The joint security architecture is intended to address deployment in sensitive environments.

The portfolio is ambitious and many claims still require production-scale validation.

Even so, it reflects the right competitive unit.

The future of domestic computing will not be determined by which company can display the largest theoretical number at an exhibition. It will be determined by which system can run valuable workloads reliably, migrate software efficiently, and remain economically viable after deployment.

常见问题

What is Tecorigin T2 Ultra?

T2 Ultra is Tecorigin’s new heterogeneous many-core processor based on the HCU architecture. It supports autonomous and accelerator modes and covers precision formats from FP64 to FP4.

What is the HCU architecture?

HCU stands for heterogeneous converged computing. It places CPU resources and AI-computing cores on a shared high-speed fabric and allows their ratio to be adjusted for different workloads.

What is HyperIntelliX?

HyperIntelliX is Tecorigin’s converged large-scale computing system for both AI and high-performance scientific computing. It is designed around T2 Ultra, liquid cooling, large-scale interconnection, and a shared software platform.

Why does T2 Ultra include FP64 support?

FP64 is important for scientific simulations and engineering workloads that require high numerical precision. Its inclusion helps the processor support AI4Science and traditional HPC rather than focusing only on low-precision AI inference.

Can Tecorigin hardware run mainstream AI frameworks?

Tecorigin’s official materials list support for frameworks and tools including PyTorch, PaddlePaddle, Megatron-LM, DeepSpeed, vLLM, SGLang, Transformers, JAX, and MindSpore. The depth of support and required migration work can vary by model and operator.

What is Tecorigin’s AI4Science platform?

It is an integrated research-computing platform that combines domestic hardware, job scheduling, development environments, programming models, scientific libraries, operator tools, and framework support for fields such as life sciences, weather, quantum simulation, fluid dynamics, and materials.

Are T2 Ultra’s performance figures independently verified?

The peak FP64, FP16, FP4, interconnect, and scaling specifications in the launch materials are vendor-reported. Independent application benchmarks and production deployment data are needed for complete comparison.

What is the full-domain large-model security system?

It is a multi-party security architecture released at WAIC
2026. It combines domestic trusted computing with model evaluation, governance, data compliance, data security, software supply-chain security, and training.

相关工具

• Tecorigin Developer Platform: Tecorigin’s official developer portal for models, ecosystem resources, and platform access.
• PyTorch: A widely used open-source framework for model training, research, and deployment.
• Megatron-LM: NVIDIA’s open-source framework for training large transformer models at scale.
• DeepSpeed: Microsoft’s open-source system for distributed model training and inference optimization.
• vLLM: A high-throughput open-source serving engine for large language models.
• JAX: A numerical-computing framework used in machine learning and scientific workloads.

Related Links

• Tecorigin Official Website: Official company, product, solution, and ecosystem information.
• Tecorigin Product and Technology Overview: Official information on the heterogeneous architecture, hardware systems, software stack, supported models, and computing management.
• Tecorigin AI4Science Platform: Official details on the domestic AI4Science platform, programming tools, supported frameworks, and research cases.
• Tecorigin Documentation Center: Technical documentation and case studies for Tecorigin products and ecosystem projects.
• Tecorigin Developer Platform: Model resources and developer ecosystem access.
• WAIC 2026 Product Report: A separate report describing T2 Ultra, HyperIntelliX, the AI4Science platform, and the large-model security system.
• National Science and Technology Innovation Center Report: A government-hosted overview of Tecorigin deployments in AI, weather, biology, and intelligent-computing centers.

Summary

Tecorigin’s WAIC 2026 launch connects a new HCU heterogeneous architecture, the T2 Ultra processor, the HyperIntelliX converged computing system, developer tools, an AI4Science platform, and a layered security system.

The strategy addresses a real change in AI infrastructure. Agent workloads require more general-purpose coordination, while scientific AI needs high-precision numerical computing, long-running jobs, and scalable parallel systems.

The architecture and vendor-reported specifications are promising, but production availability, software maturity, cluster stability, application performance, and total operating cost still require broader validation.

The main significance of Tecorigin’s launch is not one chip specification; it is the attempt to build one complete computing foundation for AI, agents, scientific computing, software migration, and secure deployment.