算力交易中心异构算力调度实录:无问芯穹接入 16 种芯片,模型供应商如何一键跨卡?

The core of this disruptive capability lies in one key technology: heterogeneous computing scheduling.


One-Click Scheduling for 16 Chip Architectures: The Breakthrough Battle of Heterogeneous Computing in Computing Power Trading Centers

At WAIC 2026, Infinicore released a striking set of industrial data: its nationwide deployed computing power has exceeded 37,000P, covering 16 mainstream chip architectures. The daily token calling volume has surged 40 times compared with December last year, marking a new milestone for large-scale heterogeneous computing commercialization in China’s AI industry.

The platform integrates 16 types of mainstream chips including NVIDIA, Huawei Ascend, Cambricon and Hygon, which differ greatly in architecture, generation and manufacturer. In the past, these heterogeneous chips operated in isolation with no interconnection or collaborative scheduling, forming countless fragmented computing silos. Now, Infinicore’s distributed computing hub unifies aggregation, elastic scheduling and on-demand distribution for all hardware resources, waking up massive idle computing potential.

The core of this disruptive capability lies in one key technology: heterogeneous computing scheduling.

Computing Fragmentation: The Underrated Core Industry Pain Point

The shortage of high-end AI computing power is a widely recognized industry dilemma, yet the more hidden and restrictive problem for enterprises is computing fragmentation.

A typical AI enterprise usually possesses a mixed batch of hardware resources: hundreds of A100 GPUs, dozens of H100 GPUs, and multiple domestic test chips. These devices are scattered across different data centers, cloud environments and procurement batches. Under traditional operation modes, hardware resources work independently — A100 for model training, H100 for inference tasks, and domestic chips left idle or only undertaking low-margin edge tasks. Isolated clusters turn massive computing resources into static, unusable fixed assets.

Infinicore completely reverses this situation by connecting fragmented computing silos and activating idle static resources into efficiently circulating dynamic assets. Relying on three self-developed core technologies — computing-communication overlapping pipeline scheduling, adaptive traffic scheduling, and full-dimensional heterogeneous resource scheduling — the platform realizes unified management and collaborative scheduling of GPU clusters scattered across different regions and cloud platforms.

This technical system has passed rigorous production-level verification in three core scenarios, achieving industry-leading performance improvements:

Ultra-long-distance cross-domain aggregation. Cooperating with China Telecom, Infinicore completed cross-domain hybrid model training spanning over 4,000 kilometers from Hami, Xinjiang to Shenzhen, Guangdong, boosting overall cluster performance by 165%. It breaks the long-standing industry consensus that ultra-long-distance WAN training is impossible due to high latency.

Multi-data-center hybrid aggregation. It connects four GPU clusters of different generations to support collaborative training of nearly billion-parameter large models, improving multi-hardware synergy efficiency by 41%. Legacy heterogeneous computing resources that were once considered cumbersome historical issues have been transformed into schedulable high-value assets.

Hybrid cloud resource interconnection. It enables secure intercommunication and hybrid training between on-premises private clusters and public cloud computing resources, lifting overall system performance by 68%. It breaks the isolated operation of enterprise local computing and cloud idle resources, maximizing resource utilization.

"One-Click Cross-Chip Scheduling": From Slogan to Industrial Engineering Implementation

The core demand of model suppliers is simple and clear: realizing one-click model deployment on the optimal hardware without manually adapting to underlying chip differences.

Infinicore solves this pain point with its pioneering cross-cluster heterogeneous PD separation architecture (PDD), bringing revolutionary upgrades to LLM inference efficiency.

LLM inference is divided into two core stages: compute-intensive Prefill and memory-access-intensive Decode. Different chips show completely differentiated performance advantages in the two stages. For example, certain domestic E-series chips only reach 81% of flagship GPU performance in the Prefill stage, but exceed 210% of flagship GPU efficiency in the Decode stage.

Traditional integrated cluster scheduling forces all chips to bear both Prefill and Decode tasks simultaneously. The differentiated performance advantages of heterogeneous chips cannot be exerted, and hardware "specialties" become limitations that drag down overall cluster efficiency.

Infinicore’s PDD architecture innovatively separates Prefill and Decode tasks, scheduling them to chips with corresponding performance advantages respectively, and interconnects homogeneous clusters in different regions via WAN Ethernet. Actual production tests show that the first token latency is reduced by 51.5%, and single-token inference cost is cut by 37.5%.

For model suppliers, underlying chip differences are completely transparent. The platform intelligently schedules compute-heavy Prefill tasks to high-computing chips and memory-heavy Decode tasks to high-bandwidth chips, realizing automatic optimal matching of tasks and hardware.

Agent Tasks: The Catalyst for Large-Scale Heterogeneous Scheduling Popularization

Xia Lixue, CEO of Infinicore, disclosed a key industry data: over 95% of inference tasks in the first half of this year originated from Agent business demands.

Agent tasks feature hierarchical decomposition — a single complex intelligent task will derive numerous sub-tasks with varying complexity. Not all sub-tasks require flagship high-end GPUs. Cost-effective domestic chips can undertake matching lightweight tasks without bearing the ultra-high load of large model training, giving full play to their marginal advantages.

Task hierarchical decomposition makes hierarchical scheduling of heterogeneous chips feasible. The significant improvement in domestic chip adaptation efficiency stems from two core factors: technological breakthroughs in cross-cluster scheduling, and the task stratification brought by Agent business explosion.

In the past year, Infinicore has reduced the overall industry inference cost to one-tenth of the original level. According to Xia Lixue’s prediction, there is still 90% cost reduction space under full-link collaborative optimization in the future.

Unexhub: Completing the Final Puzzle of the Computing Power Trading Ecosystem

Infinicore’s heterogeneous scheduling solves the core problem of how to maximize computing resource utilization, while Unexhub solves the key problem of how to realize efficient circulation and monetization of computing resources. The two complement each other to build a closed-loop industrial ecosystem.

Unexhub comprehensively reshapes the global token distribution system and builds a leading aggregated token trading and distribution center, with three core mechanisms to empower suppliers and developers:

Real-time Settlement. It eliminates long account periods and sluggish capital recovery in traditional industries. The professional financial clearing system enables suppliers to withdraw funds in real time based on actual resource supply, completely zeroing out financial pressure and converting idle GPU cluster assets into sustainable cash flow.

Early Zero-Commission Incentive. Exclusive benefits for early settled suppliers with permanent zero-platform-commission access, drastically lowering the entry threshold for small and medium-sized computing and model suppliers.

Dynamic API Bidding. API calling prices float intelligently with real-time market supply and demand. Developers enjoy ultra-low-cost resources during off-peak hours, while suppliers gain higher profit margins during peak demand, realizing a win-win for both sides.

Two Paths, One Industrial Essence: Turn Static Resources Into Dynamic Assets

If Infinicore’s computing distributed hub is regarded as the "production workshop" of tokens — efficiently activating and collaborating 16 types of heterogeneous chips to produce high-value token resources, Unexhub is the open "circulation market" of tokens — realizing free trading, dynamic pricing and instant monetization of computing assets.

Infinicore improves production efficiency through heterogeneous scheduling technology; Unexhub accelerates circulation efficiency through financial trading mechanisms.

With one-click intelligent scheduling for 16-chip heterogeneous clusters and full real-time settlement of token transactions, the industry’s core competition logic has completely shifted. Model suppliers no longer struggle with "whether there is sufficient computing power", but focus on "how to use computing resources at the lowest cost and highest efficiency".

This is the essential value of modern computing power trading centers: not simply expanding resource pipelines, but promoting the overall upgrading of the computing industry from "planned resource allocation" to "market-oriented free circulation".

Unexhub Platform Core Strengths & Exclusive Benefits

Unexhub has gathered a full spectrum of global AI suppliers, covering top-tier large models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality niche models from SMEs and individual developers. Relying on millisecond-level intelligent routing technology, massive AI resources participate in open market bidding, delivering ultra-low-cost and ultra-stable calling services for enterprise users and developers.

Exclusive User Benefits: Developers pay strictly based on actual token consumption, with the lowest discount reaching 30% of the market official price. Suppliers can access the platform with one click and enjoy real-time settlement of calling revenue, realizing zero-threshold monetization of idle computing and model resources.


16种芯片一键调度:算力交易中心的异构计算突围之战

2026年7月世界人工智能大会(WAIC)上,无问芯穹发布一组重磅行业数据:平台全国部署算力规模突破37000P,全面覆盖16类市面主流AI芯片,Token日调用量较去年12月暴涨40倍,标志着国内异构算力规模化商用迈入全新阶段。

平台汇聚的16种主流芯片,涵盖英伟达、华为昇腾、寒武纪、海光等全品类硬件,囊括不同架构、不同代际、不同厂商的异构算力资源。过往,各类芯片各自独立运行、无法互联互通,形成大量算力孤岛与资源闲置;如今,无问芯穹「算力集散中心」实现全品类芯片统一汇聚、弹性调度、按需分发,彻底盘活行业存量算力资源。

这套打破行业壁垒的核心能力,凝练为三个字:异构调度。

算力碎片化:被行业长期低估的核心痛点

行业普遍聚焦高端算力紧缺问题,却忽略了更制约企业降本增效的隐形痛点:算力碎片化。

多数AI企业的硬件资源均处于零散分布状态:既有数百张A100、数十张H100高端显卡,也搭配各类国产测试芯片,设备分散在不同数据中心、不同云平台、不同采购批次中。传统运维模式下,各类芯片各司其职、互不协同:高端显卡专属模型训练,中端显卡承载常规推理,国产芯片长期闲置或仅承接边缘轻量化任务。算力集群彻底固化为孤岛,海量硬件沦为无法复用的「死资源」。

无问芯穹的核心解法,是打通所有算力孤岛,让静态闲置资源流转复用。依托自研三大核心技术——计算通信重叠流水编排、自适应通信流调度、全维度异构算力调度,平台实现跨地域、跨云环境、跨架构GPU集群的统一管控与协同算力。

这套自研技术体系已通过三大生产级场景严苛验证,落地效果远超行业预期:

超远距离跨域聚合 联合中国电信完成新疆哈密至广东深圳、跨度超4000公里的广域网跨域混训,集群整体性能提升165%,彻底推翻「跨千里网络无法协同训练、延迟过高不可用」的行业固有认知。

多数据中心跨代聚合 打通4个不同代际的异构GPU集群,支撑近百亿参数大模型联合训练,资源协同效率提升41%。过往被视为累赘的「历史遗留异构算力」,成功转化为可高效调度的优质资产。

混合云公私互通聚合 实现企业本地私有集群与公有云算力的安全互通、混合训练,整体算力性能提升68%,打破本地算力与云端闲置资源割裂使用的行业现状,最大化释放资源价值。

「跨卡一键调度」:从行业口号落地为硬核工程能力

所有模型供应商的核心诉求高度统一:无需适配底层芯片差异,一键将模型部署至最优硬件载体,实现极致性能与最低成本的平衡。

无问芯穹首创跨集群异构PD分离架构(PDD),精准破解行业适配难题,实现真正意义上的异构芯片透明化调度。

大模型推理分为两大核心阶段:计算密集型的预填充(Prefill)、访存密集型的解码(Decode)。不同架构芯片在两大阶段性能差异极大:部分国产E系列芯片,Prefill阶段性能仅为旗舰GPU的81%,但Decode阶段性能可飙升至旗舰GPU的210%,具备极强的场景化优势。

传统集群调度模式存在致命短板:所有芯片统一承载Prefill与Decode全流程任务,无法发挥异构芯片的差异化优势,硬件「长板能力」被限制,短板反而拖累整体集群推理效率。

无问芯穹PDD架构创新性实现两段任务拆分,将Prefill、Decode任务分别调度至对应优势芯片,再通过广域网以太网串联跨地域同构集群。实测数据显示:模型首Token延迟降低51.5%,单Token推理成本直接下降37.5%,降本增效效果极其显著。

对模型供应商而言,底层芯片架构、厂商、代际差异完全无感。平台智能匹配任务与硬件,算力型芯片承载高负载预填充任务,访存型芯片承接高频率解码任务,全程无需人工干预,真正实现异构算力透明化最优调度。

Agent任务爆发,成为异构调度普及的核心催化剂

无问芯穹CEO夏立雪披露关键行业数据:2026年上半年,平台超95%的推理任务均来自Agent智能体业务。

Agent任务具备典型的分层拆解特性:一项复杂智能任务,会衍生出大量复杂度不同、算力需求各异的子任务,并非所有任务都需要高端旗舰芯片支撑。高性价比国产芯片可精准承接适配自身性能的轻量化任务,无需强行承载大模型超高负载,最大化释放国产算力价值。

任务分层特性,让异构芯片分层调度成为行业刚需,也是本轮国产芯片适配效率大幅提升的核心原因。一方面是跨集群异构调度技术实现突破,另一方面是Agent业务爆发带来的任务结构革新,双向驱动异构算力规模化落地。

过去一年,无问芯穹通过技术迭代,将行业整体推理成本压缩至原先的十分之一。夏立雪预判,随着全链路协同优化持续落地,行业算力成本未来仍有90%的下降空间。

Unexhub:补齐算力交易生态最后一块拼图

如果说无问芯穹的异构调度,解决了「算力如何高效利用」的生产端问题;那么Unexhub则补齐了「算力如何高效流转变现」的流通端问题,二者相辅相成,构建完整算力产业闭环。

Unexhub重构全球Token分发底层逻辑,搭建全球化词元聚合交易中心,依托三大核心机制,全方位赋能供需两端,激活行业算力资产活力:

即时清算结算 彻底解决传统算力供应商账期长、回款慢、现金流承压的痛点,平台专属金融清算体系,支持供应商按实时供给量全额回款、即时提现,清零中小团队财务压力,让闲置GPU硬件资产转化为稳定现金流。

早期零佣金激励 面向早期入驻供应商开放永久零平台佣金权益,大幅降低行业入驻门槛,中小算力、模型供应商可零成本入局、零门槛变现。

API动态竞价机制 打破平台固定定价垄断,API调用价格随市场供需实时浮动。开发者闲时可薅取低价算力、极致降本;供应商忙时可抬升收益、最大化资源价值,实现供需双向共赢。

两条赛道殊途同归:让算力从「静态资源」变身「流动资产」

无问芯穹算力集散中心,是AI词元的「生产车间」:通过16种芯片异构协同调度,盘活全域零散算力,高效生产优质Token资源,提升行业生产效率;

Unexhub算力交易中心,是AI词元的「流通市场」:通过市场化、金融化交易机制,让Token算力自由流通、动态定价、即时变现,加速行业流通效率。

异构调度优化生产效率,金融交易加速流通效率,两大核心能力双向赋能,彻底改写行业规则。

当16类异构芯片实现一键智能调度、Token算力实现实时竞价结算,行业核心矛盾彻底转移:模型供应商不再纠结「有没有算力」,而是聚焦「怎么用算力最划算、变现最高效」。

这正是算力交易中心的终极价值:不止是搭建算力流通管道,更是推动整个AI算力行业,从「计划式资源分配」全面迈入「市场化自由流通」新时代。

Unexhub平台核心权益,普惠全行业从业者

目前Unexhub已汇聚全球全量级AI供应商,覆盖Gemini、ChatGPT、Claude、DeepSeek、智谱GLM等头部顶级大模型,同时吸纳海量中小企业、个人开发者的优质特色模型。依托毫秒级智能路由技术,全网AI资源公开竞价流通,为用户提供极致低价、极速稳定的调用服务。

平台硬核专属福利:开发者全程按Token实际消耗计费,最低可享官方市场价3折极致低价;供应商一键入驻开通变现渠道,调用分成实时结算,零门槛盘活闲置AI算力资产,轻松实现稳定盈利。