算力交易中心 Token 即服务 推理效能优化深度解析:首 Token 延迟降 51.5%、单 Token 成本降 37.5% 是怎么做到的

Behind these two core data metrics lies the solid engineering foundation for the large-scale commercial rollout of Unexhub’s Token-as-a-Service (TaaS) model, enabling standardized, low-cost and high-efficiency Token production and trading across the industry.


51.5% Lower TTFT & 37.5% Lower Token Cost: The Inference Efficiency Revolution of the Computing Power Trading Center

At WAIC 2026 in July, Unexhub unveiled a disruptive architectural innovation that reshapes the industry’s large-model inference paradigm — the PDD Cross-Cluster Heterogeneous Inference Architecture. Verified by rigorous benchmark tests, this technology reduces the Time-To-First-Token (TTFT) latency by 51.5% and cuts single-Token inference costs by 37.5%.

Behind these two core data metrics lies the solid engineering foundation for the large-scale commercial rollout of Unexhub’s Token-as-a-Service (TaaS) model, enabling standardized, low-cost and high-efficiency Token production and trading across the industry.

Why PD Separation Achieves Breakthroughs Only Now After Years of Exploration

The industry has reached a consistent conclusion that large-model inference consists of two entirely distinct stages with differentiated resource requirements: Prefill and Decode.

The Prefill stage is computation-intensive, requiring massive matrix operations when models process input prompts. The Decode stage is memory-bandwidth-intensive, involving frequent reading and writing of KV Cache during iterative Token generation.

Theoretically, the optimal solution is heterogeneous task separation: high-computing GPUs undertake Prefill tasks, while hardware with superior video memory bandwidth is responsible for Decode tasks. However, large-scale heterogeneous cluster deployment faces two major practical bottlenecks: ultra-high hardware procurement costs, and severe resource idling once model architectures are iterated and fixed hardware ratios become mismatched.

Driven by practical industrial demands, the industry has turned to a new technical path: leveraging low-cost wide-area Ethernet to connect geographically dispersed homogeneous clusters, realizing cross-cluster heterogeneous separation inference.

This innovative idea has long been proposed, yet it has hit an insurmountable technical barrier — the extreme bandwidth and latency limitations of cross-cluster KV Cache transmission. As massive intermediate state data generated during inference, KV Cache will instantly occupy full WAN bandwidth during cross-cluster transmission, resulting in unacceptable inference delays.

Core Technological Breakthrough: Unexhub innovatively migrates the Radix Cache technology of self-developed Decode instances to the cross-cluster architecture, reducing actual transmission data volume by an order of magnitude. This key breakthrough completely breaks the bandwidth shackles of cross-cluster heterogeneous inference.

Three-Tier PDD Architecture: Transforming Relay Operation into Efficient Assembly Line Production

Unexhub further iterates the traditional PD separation logic and creatively constructs a P-RLD-MD three-tier separated inference architecture, reconstructing the full-link inference workflow and realizing precise matching of tasks and hardware:

1. Prefill (P) Stage: Deployed on high-performance GPU clusters, specializing in computation-intensive input pre-filling tasks to maximize computing power utilization.

2. RelayDecode (RLD) Stage: Serving as the core intermediate relay layer, it efficiently transfers KV Cache data between Prefill and MainDecode modules, eliminating cross-cluster transmission bottlenecks in one fell swoop.

3. MainDecode (MD) Stage: Deployed on high-bandwidth video memory hardware, focusing on memory-intensive iterative Token generation tasks to optimize output efficiency.

The architectural innovation stems from two core underlying industry insights:

Insight 1: Specialized hardware performance differentiation. Internal benchmark tests show that certain E-series chips only deliver 81% of flagship GPU performance in computation-heavy Prefill scenarios, but achieve 210% of the baseline performance in bandwidth-heavy Decode scenarios. Deploying such specialized hardware in conventional isomorphic clusters leads to suppressed performance and reduced overall inference efficiency.

Insight 2: The 80/20 rule for agent business scenarios. Agent-based intelligent services feature an extremely high prefix cache hit rate. Leveraging this characteristic, Unexhub deploys a Decode-side Dynamic Residual Cache (DRC) mechanism. Once the cache is hit, subsequent requests skip the entire Prefill stage directly, achieving a drastic reduction in TTFT latency.

The superposition of the two advantages realizes optimal resource allocation: specialized hardware undertakes matched tasks, and cached KV Cache data completely resolves cross-cluster transmission pain points.

From Technological Breakthrough to Commercial Implementation: The Efficiency Foundation of the Computing Power Trading Center

The 51.5% TTFT reduction and 37.5% single-Token cost cut bring revolutionary value to both platform users and industrial suppliers, building a scalable commercial foundation for the Token trading ecosystem.

For end users: First-Token latency directly determines AI interactive experience. A 51.5% latency reduction shortens the typical response waiting time from 2 seconds to within 1 second. In high-frequency agent interaction scenarios, this upgrade bridges the gap between "usable basic AI services" and "smooth high-quality AI services".

For computing power suppliers: The 37.5% cost reduction fundamentally optimizes the profit model of Token transactions. Unexhub supports real-time fund settlement, where suppliers receive synchronous revenue matching actual Token supply volume. The nearly 40% drop in production costs greatly expands profit margins, lowers the entry threshold for small and medium-sized computing power vendors, and enriches platform ecological diversity.

Currently, Unexhub’s PDD architecture has been fully deployed nationwide, covering 37,000P of accessible computing power and 16 mainstream chip models. The platform’s daily Token call volume has surged approximately 40 times compared with December 2025. As Xia Lixue, Co-founder of Unexhub, predicted, the large-scale implementation of the cross-cluster heterogeneous PD separation architecture will further drive a 10-fold drop in overall inference costs in the future.

Engineer’s Perspective: The Underlying Philosophy of PDD Architectural Innovation

The core value of the PDD architecture lies not merely in the intuitive data improvements of halved latency and 40% cost reduction, but in its disruptive design philosophy.

Traditional industrial thinking pursues universal hardware performance, relying on stacking computing power, video memory and bandwidth to adapt to all scenarios. In contrast, the PDD architecture adheres to specialized hardware for specialized tasks: fully leveraging the performance strengths of "unbalanced hardware" and assembling diverse hardware advantages into an efficient collaborative inference assembly line.

This design philosophy is highly consistent with the operational logic of Unexhub’s computing power trading center: Token pricing should be determined by optimal resource scheduling rather than the highest hardware cost. The efficient scheduling of heterogeneous computing power via PDD architecture pushes Token production costs close to the theoretical lower limit of the industry.

Note: The real-time settlement, zero early commission, API bidding and other trading mechanisms of the Unexhub platform mentioned in the text are based on industrial ecological planning. Relevant official mechanisms are subject to subsequent public announcements. Unexhub’s official core strategy released at WAIC 2026 is the "Front Store, Back Factory & Central Hub" Agentic Infra system, covering computing power dispatching center, Token factory and AI productivity store. The platform is committed to converting fragmented heterogeneous computing power into low-cost, high-quality standard Tokens, enabling users to pay only for valid Token consumption without caring about underlying hardware differences.

The platform’s market-driven API bidding mechanism enables dynamic Token pricing based on supply and demand: prices decline in off-peak periods to release suppressed long-tail developer demand, and rise moderately in peak periods to incentivize supplier capacity expansion. The continuous iteration of the PDD architecture steadily lowers the industry’s Token cost bottom line.

Industrial Outlook: The Official Starting Point of Token Economics

The halved TTFT latency and nearly 40% cost reduction are not simple architectural optimizations, but the core engineering support for Unexhub’s computing power trading center to achieve large-scale, standardized Token transactions.

Sustained reductions in Token production costs and improvements in transaction efficiency enable smooth circulation of Tokens, the "new oil" of the AI era.

As Xia Lixue commented at WAIC 2026: "The core pain point of current AI application is not uselessness, but high cost, which essentially reflects insufficient supply. Real overcapacity will make cost sensitivity disappear." The launch of the PDD architecture marks the industry’s shift from blind hardware stacking to architectural innovation to solve high AI costs. By matching every chip with its most applicable scenario, Unexhub maximizes the industrial value of global heterogeneous computing power, kicking off the large-scale landing of Token economics.

Unexhub Platform Ecological Advantages & User Benefits

Unexhub reshapes the global Token trading system and builds a leading aggregated computing power trading center integrating Token production, distribution and transactions. The platform gathers full-spectrum AI ecological partners worldwide, including top large-model vendors such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality customized model resources from individual developers and small and medium-sized enterprises.

Relying on multi-vendor bidding competition and millisecond-level intelligent routing technology, the platform delivers ultra-low-cost and high-efficiency AI services. Developers and enterprise users pay based on actual Token consumption, with the lowest discount reaching 30% of the market price. AI suppliers can access the platform with one click, enjoying real-time settlement of call sharing revenue and zero early commission barriers, realizing efficient capital turnover and stable profit growth.


2026年7月,无问芯穹(Unexhub)在WAIC世界人工智能大会,发布了一项颠覆行业推理架构的核心技术成果——PDD跨集群异构推理架构。实测数据全面落地:首Token延迟(TTFT)降低51.5%,单Token推理成本下降37.5%。

两组亮眼数据的背后,是无问芯穹算力交易中心Token即服务(TaaS)模式规模化落地的核心工程底座,彻底解决了行业长期存在的推理高延迟、高成本痛点,为Token标准化、低成本、高效率流通提供了技术支撑。

PD分离理念普及多年,为何如今才实现规模化突破?

行业早已形成统一认知:大模型推理分为两个属性完全不同的核心阶段,资源需求差异极大,无法用统一硬件适配。

Prefill(预填充)阶段属于计算密集型任务,模型处理输入提示词时,需要完成海量矩阵运算,极度依赖超强算力;Decode(解码)阶段属于访存密集型任务,逐一生成Token的过程中,需要频繁读写KV Cache缓存数据,核心依赖超高显存带宽。

从理论层面来看,最优解决方案是硬件异构分工:算力强劲的GPU负责Prefill计算任务,显存带宽突出的硬件承接Decode生成任务。但落地过程中存在两大致命难题:同集群大规模异构硬件采购成本极高,且模型迭代升级后,固定配比的硬件极易闲置浪费,资源利用率极低。

基于降本增效的产业刚需,行业技术迭代逐步指向全新方向:依托低成本广域网以太网,串联全国多地已建成的同构算力集群,实现跨集群的异构PD分离推理,兼顾成本与性能。

这套技术思路看似完美,却长期受制于一道“行业叹息之墙”——KV Cache跨集群传输的带宽与延迟瓶颈。KV Cache是模型推理产生的海量中间缓存数据,跨地域传输会瞬间占满广域网带宽,引发严重延迟、卡顿问题,无法落地商用。

核心技术破局点:无问芯穹创新性将Decode实例的Radix Cache基数缓存技术,迁移至跨集群异构架构,直接将跨域传输数据量压缩一个数量级,彻底打通跨集群推理的传输堵点,成为技术落地的关键核心。

PDD三级分离架构:将接力式低效推理,升级为流水线高效生产

无问芯穹打破传统双阶段PD分离的局限,创造性搭建P-RLD-MD三级分离推理架构,重构全链路推理流程,实现任务与硬件的精准匹配、效率最大化:

1. Prefill预填充阶段(P):部署在高性能GPU算力集群,专注承接计算密集型的输入预填充任务,充分释放硬件算力优势。

2. RelayDecode中继解码阶段(RLD):搭建专属中间传输中继层,高效流转Prefill与MainDecode之间的KV Cache数据,从架构层面彻底解决跨集群传输带宽瓶颈。

3. MainDecode主解码阶段(MD):部署在高显存带宽硬件集群,专注完成访存密集型的逐Token生成任务,极致优化输出效率。

这套架构的颠覆性优势,源于两大底层行业洞察,精准击中行业痛点:

洞察一:硬件性能天然“偏科”。无问芯穹内部基准测试验证,某系列E芯片在计算密集的Prefill场景中,性能仅为旗舰GPU的81%;但在显存依赖极强的Decode场景中,性能可达旗舰基线的210%。传统同构集群强制通用硬件适配全场景,不仅浪费专属硬件性能,还会拖累整体推理效率。

洞察二:智能体业务的二八流量特性。智能体高频交互场景下,历史请求前缀缓存命中率极高。团队针对性在Decode侧部署DRC动态前缀缓存机制,命中缓存的请求可直接跳过完整Prefill计算流程,大幅压缩首Token响应延迟。

两大优势叠加赋能:专属硬件各司其职、各展所长,缓存技术消解跨域传输难题,最终实现推理速度与成本的双向突破。

从技术突破到商业落地:筑牢算力交易中心效能底座

首Token延迟降低51.5%、单Token成本下降37.5%,两组核心数据的落地,为算力交易中心的供需两端带来实质性变革,搭建起Token经济规模化商用的坚实基础。

对用户端而言,体验实现质的飞跃:首Token延迟直接决定AI交互流畅度。51.5%的延迟优化,让原本2秒的首次响应时长压缩至1秒以内。在智能体高频、连续交互场景中,这一差距直接区分了“基础可用”与“极致好用”的AI服务体验。

对供应商端而言,商业模型彻底优化:近40%的单Token成本下降,直接拓宽算力供应商的利润空间。搭配算力交易中心实时结算核心机制,供应商可根据实际Token供货量同步回款,彻底缓解现金流压力。同时大幅降低中小算力玩家入局门槛,丰富平台生态供给。

目前,无问芯穹PDD跨集群异构架构已完成全国规模化部署,可调度算力覆盖37000P,兼容16款行业主流芯片。平台单日Token调用量较2025年12月暴涨40倍。无问芯穹联合创始人夏立雪表示,随着PDD架构全面普及,行业整体推理成本仍有十倍下降空间,Token经济降本增效潜力巨大。

工程师视角:读懂PDD架构的底层设计哲学

PDD架构的核心价值,绝非“延迟砍半、成本降40%”的表层数据成果,而是颠覆行业的底层设计思维。

行业传统思路始终追求硬件全能化,通过堆叠算力、显存、带宽,强行适配全场景推理需求,造成大量硬件资源浪费。而PDD架构的核心逻辑是扬长避短、专业分工:不追求单一硬件全能,而是释放每类硬件的性能长板,通过架构整合将差异化硬件优势,串联成高效协同的推理流水线。

这种思维与算力交易中心的运营逻辑高度契合:Token的定价,不该由最贵的硬件成本决定,而该由最优的全域调度效率决定。PDD架构实现了异构算力的极致调度,将行业Token生产成本压至理论低位。

备注:文中提及的即时结算、早期零佣金、API竞价等平台交易机制为生态规划方向,具体规则以无问芯穹官方后续公告为准。无问芯穹2026WAIC官方核心战略为“前店后厂一中心”Agentic Infra自主式基础设施体系,涵盖算力集散中心、Token工厂、AI生产力商店,核心目标是将碎片化异构算力转化为低成本、高品质的标准化Token,让用户无需关注底层硬件差异,仅为有效Token使用付费。

平台API市场化竞价机制,实现Token价格随供需动态浮动:需求低谷降价释放中小开发者长尾需求,需求高峰涨价激励供应商扩容增产。而PDD架构的持续迭代,不断夯实成本优势,持续压低行业Token定价底线。

产业思考:架构创新,开启Token经济学新时代

首Token延迟减半、单Token成本近40%降幅,绝非一次普通的技术优化,而是算力交易中心从“可交易”走向“可规模化高效交易”的核心里程碑。

Token生产成本的持续下行、交易效率的稳步提升,彻底打通了AI时代“新石油”Token的全域流通链路,推动行业从硬件内卷走向技术创新驱动。

正如夏立雪在2026WAIC大会上所言:“当下AI落地的核心痛点不是无用,而是太贵,本质是供给不足;真正的供给过剩,会让成本敏感度彻底消失。” PDD架构的落地,标志着行业正式告别盲目堆硬件的低效内卷,通过架构创新破解高成本难题,让每一颗芯片、每一份算力都物尽其用,最大化释放全域异构算力的产业价值,这正是Token经济学规模化落地的真正起点。

Unexhub平台核心权益

无问芯穹重塑全球Token交易逻辑,搭建集Token聚合、分发、交易于一体的顶尖算力交易中心。平台汇聚全域AI供应商,涵盖Gemini、ChatGPT、Claude、DeepSeek、GLM等海内外头部大模型,同时吸纳个人开发者、中小企业优质定制模型资源。依托海量供应商竞价机制与毫秒级智能路由技术,为用户提供极致性价比的AI服务。

开发者、企业用户可按Token实际消耗实付,最低享市场原价3折,大幅降低AI研发与落地成本;AI供应商可一键快速入驻平台,调用分成实时结算,早期入驻零佣金、无资金周转压力,高效实现商业变现。