以场景定义架构 隼瞻科技发布Axon Platform全定制AI算力平台Scenario-Defined Architecture: FZC Tech Launches Axon Platform Fully Customized AI Computing Platform


AI free token/AI免费积分
0

10月20日,专注RISC-V处理器架构与AI算力平台研发的科创企业——隼瞻科技(广州)有限公司(以下简称:隼瞻科技)主办的"Axon Platform AI算力平台发布会"在深圳举行,发布会主题"架构级定制·为AI情形而生"。在AI应用向垂直行业深度渗透、通用GPU在功耗能效与成本上遭遇瓶颈、RISC-V开放底座加速崛起的背景下,Axon Platform正是为填补“面向AI情形的架构级定制”这一空白而生。

以场景定义架构 隼瞻科技发布Axon Platform全定制AI算力平台

发布会上,隼瞻科技CEO曾轶、CTO姚彦斌、南方科技大学助理教授兼隼瞻科技联合研究中心AI研究科学家林龙扬、隼瞻机型市面总监陆斌先后登台,浦发银行华润城支行对公客户经理罗一鹏作科技金融做相关主题分享。各IC外观企业负责人、高院校代表、投资机构与半导体媒体代表等产业人士齐聚一堂,共同见证AI算力"为情形而生"的新起点。

"为情形而生"是方法,不是口号

发布会伊始,隼瞻科技CEO曾轶致欢迎辞。他用一组极其接地气的情形描绘了行业现状:“以前人们总说‘AI 是在为 Token 打工’,云端算力成本居高不下。如果笔者能让一台千元级别的打印机塞进一个专用语言模型,达成端侧商业模式从To B到To C的进化,那才是真正的行业裂变。”曾轶认为,从 2025 年端侧AI 元年到 2026 年新形态硬件涌现,AI 算力正从云端走向本地,安全隐私、毫秒级响应和成本倒逼,让“一情形一架构”成为必然。

以场景定义架构 隼瞻科技发布Axon Platform全定制AI算力平台

隼瞻科技CEO曾轶

“AI应用的多样性与端侧芯片在功耗、面积上的约束,是一对永恒的矛盾。”曾轶指出,“以前是芯片公司决定架构,现在架构的决定权跑到了AI公司那里。”

随后,曾轶以《为情形而生的AI算力平台:Axon Platform》为演讲主题正式发布Axon Platform。他用“太极的阴阳两面”来比喻这个平台:硬件架构是“阳面”,软件是“阴面”。阳面以 RISC-V 为基座,集成 Scalar 标量核、Vector矢量单元、双源张量 NPU 阵列与 NoC 互联网络,算力覆盖 0.1T 到 128T;阴面则是以 ArchStudio、Archit.ai 编译器等构成的软件平台。

在架构层面,Axon Platform把Agent AI调度任务编排做进RISC-V AI调度指令,让跨引擎调度硬件化——任务下发延迟低于2微秒,跨引擎调度开销降低约80%;张量侧使用数据流驱动架构,编译期即完成算子调度、SRAM驻留与张量Tile映射,片上数据复用率超85%,片外访存降低60%至75%,能效提高1.8至2.5倍。曾轶分享了与国际标杆企业的比较测试:“在同样体系下,笔者做出来的架构往往只有同品类思路 1/3 的面积和 1/3 的功耗。”在特定的应用情形下,Axon Platform在张量侧给出了基于SRAM存算一体工艺构建的NPU,可以进一步的降低应用的功耗。

在工具层面,Axon Platform配套ArchitStudio DSA工具、Archit.ai架构探索引擎与HiCon NoC规格工具,把"可定制"从概念变成可交付的工程能力。当下,Axon Platform已覆盖智能机器人、工业视觉、车载感知、智能安防等多类情形,既满足轻量实时推理,也能支撑高算力密集型AI任务。

把架构外观从"手艺"变成"工程"

姚彦斌以《Axon的"架构师":ArchitStudio和架构探索工具》为题,回答平台"怎么定义出来"。他用一个形象的比喻解释了痛点:“找病灶就像拍CT。”过去,端侧算力架构面临算子之殇与瀑布式开发的长周期矛盾,架构师往往靠“拍脑门”凭经验定思路。

以场景定义架构 隼瞻科技发布Axon Platform全定制AI算力平台

隼瞻CTO姚彦斌

作为 Axon 体系里赋能架构师的核心底座,ArchitStudio 给出了一套 “乐高积木式” 的搭建工具。ArchitStudio 以自研RISCAL可重定向指令集架构语言为地基,将指令的二进制编码、汇编语法、语义与流水级次定义在同一处,编译器、汇编器、仿真器、调试器全部由此自动生成;其编译器语义自动匹配工艺,让新增指令无需改动上层模型代码即可被编译器自动调度,达成 "定义即交付"。姚彦斌表示:“过去需要 3 个工程师花 5 个月完成的工作,现在第一个月思路就能落地,架构探索周期由月级缩短至周级。”

针对端侧 AI 架构外观空间巨大、维度强耦合、试错代价高的难题,Axon给出一套架构探索平台,借助"建模—搜索—PPA 联合评估"流程自动遍历候选架构,从帕累托前沿选出最优解。以 Llama 3 8B 为例,单次架构敏感性分析仅需约 17 分钟;核心周期即缩短约 4 倍,让芯片企业在流片前就能量化验证架构决策,达成从"猜架构"到"算架构"的转变。

算力引擎:全栈存算一体工艺

南方科技大学助理教授、隼瞻科技AI研究科学家林龙扬以《Axon 的“算力引擎”:全栈存算一体工艺》为题发表演讲。他将存算一体形象地比作“把存储和计算的距离缩短”,从板级、封装一路缩短到 Macro 甚至 Bitcell 级别,每缩短一个量级,能效和带宽就能提高一个量级。

以场景定义架构 隼瞻科技发布Axon Platform全定制AI算力平台

南方科技大学助理教授、隼瞻科技AI研究科学家林龙扬

林龙扬详细拆解了存算一体的四种范式(数字 vs 模拟;易失 vs 非易失)。针对精度与工艺痛点,团队分别在SRAM和MRAM验证了数字域存内计算宏,达成精度无损计算,且能效大于 100 TOPS/W。

他强调:“光靠一个存算阵列无法跑完端到端模型,阵列的峰值能效和平台能效之间有很大落差。”为此,Axon 平台给出全栈协同优化,借助 AdaPIM 存算芯片的可重构 4D-CIM 引擎与专用编译器,兼容 CNN、Transformer 及混合模型,硬件利用率最高达 83.12%。

可规格NoC决定AI ASIC的上限

陆斌以《Axon的"高速公路":可规格Hicon-NoC》为题,把话题从算力引向数据搬运。"算力决定上限,搬运决定下限。"陆斌在分享中,将NoC比作AI SoC的"高速公路"。传统共享总线在多点并发时带宽塌陷、点到点直连带来连线与面积爆炸、跨层解耦导致带宽无法连续——三重困境之下,可规格NoC成为AI ASIC破局的关键。

以场景定义架构 隼瞻科技发布Axon Platform全定制AI算力平台

隼瞻机型市面总监陆斌

Hicon-NoC 给出了路由、网络、协议接口三大可规格维度。配套 HiGUI 与 HiGen 工具链,NoC 规格迭代周期由天级缩短至 2 小时,一次借助率超 99%。量化指标同样亮眼:片上峰值带宽 256GB/s,跨核访问延迟低于 20 纳秒。在 12 TOPS 的端侧思路中,Hicon-NoC 以 4x4 Mesh 骨干互联 NPU 阵列、CPU 与存储,让平台真正跑起来。

科创赋能 金融推动

浦发银行介绍了浦发银行特色科技金融机型。科技创新的落地,离不开金融活水的精准滴灌;金融与科技的深度融合,将推动AI算力产业加速腾飞。

以场景定义架构 隼瞻科技发布Axon Platform全定制AI算力平台

浦发银行益田支行对公客户经理罗一鹏

当AI算力从"训练军备竞赛"转向"推理规模化落地",谁能把情形变成架构、把架构变成量产,谁就能在"架构主导权"的争夺中占据身位。以情形定义处理器架构,隼瞻科技用Axon Platform平台给出了自己的答案,也为AI算力产业的"架构级定制"补上了关键一块拼图——现在,平台已经搭好,接下来要看的是客户与投片给出的答案。

本文来源:https://www.163.com/dy/article/LADDUHNA0511BFPT.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

0

On October 20, Axon Platform AI Computing Platform Launch Event, hosted by FalconRise Technology (Guangzhou) Co., Ltd. (hereinafter: FalconRise Technology), an innovative enterprise focused on RISC-V processor architecture and AI computing platform R&D, was held in Shenzhen under the theme “Architecture-Level Customization · Born for AI Scenarios.” Against the backdrop of AI applications deeply penetrating vertical industries, general-purpose GPUs hitting bottlenecks in power consumption, energy efficiency and cost, and the RISC-V open foundation rising rapidly, Axon Platform was born to fill the gap of “architecture-level customization for AI scenarios.”

Scenario-Defined Architecture: FZC Tech Launches Axon Platform Fully Customized AI Computing Platform

At the launch event, FalconRise Technology CEO Zeng Yi, CTO Yao Yanbin, Lin Longyang (Assistant Professor at Southern University of Science and Technology and AI Research Scientist at the FalconRise joint research center), and FalconRise market director Lu Bin took the stage in succession. Luo Yipeng, a corporate client manager at SPD Bank Huarun City Branch, gave a tech-finance themed sharing. Industry stakeholders including IC product leaders, university representatives, investment institutions and semiconductor media gathered to witness the new starting point of AI computing power “born for scenarios.”

“Born for scenarios” is a method, not a slogan

At the start of the event, FalconRise Technology CEO Zeng Yi delivered a welcome address. He described the industry’s status quo with a set of down-to-earth scenarios: “In the past people always said ‘AI is working for Tokens,’ with cloud computing costs remaining high. If one can fit a dedicated language model into a printer at the thousand-yuan level, achieving an end-side business model evolution from To B to To C, that would be a true industry disruption.” Zeng Yi believes that from the first year of edge AI in 2025 to the emergence of new-form hardware in 2026, AI computing power is moving from the cloud to the local side; security and privacy, millisecond-level response, and cost pressures make “one scenario, one architecture” inevitable.

Scenario-Defined Architecture: FZC Tech Launches Axon Platform Fully Customized AI Computing Platform

FalconRise Technology CEO Zeng Yi

“The diversity of AI applications and the constraints of edge chips in power consumption and area are an eternal contradiction,” Zeng Yi noted. “In the past chip companies decided the architecture; now the decision-making power for architecture has shifted to AI companies.”

Subsequently, Zeng Yi officially unveiled Axon Platform with the speech topic “AI Computing Platform Born for Scenarios: Axon Platform.” He used “the two sides of the taiji (yin and yang)” to describe the platform: the hardware architecture is the “yang side,” and the software is the “yin side.” The yang side is built on RISC-V, integrating Scalar cores, Vector units, a dual-source tensor NPU array and a NoC interconnect network, with computing power covering 0.1T to 128T; the yin side is a software platform composed of ArchStudio and the Archit.ai compiler.

At the architecture level, Axon Platform embeds Agent AI scheduling and task orchestration into RISC-V AI scheduling instructions, making cross-engine scheduling hardware-based—task dispatch latency is below 2 microseconds, and cross-engine scheduling overhead is reduced by about 80%; on the tensor side, it adopts a dataflow-driven architecture, completing operator scheduling, SRAM residency and tensor tile mapping at compile time, with on-chip data reuse exceeding 85%, off-chip memory access reduced by 60% to 75%, and energy efficiency improved by 1.8 to 2.5 times. Zeng Yi shared comparative tests with international benchmark companies: “Under the same system, the architecture we build typically uses only one-third of the area and one-third of the power of the same-category approach.” In specific application scenarios, Axon Platform provides a tensor-side NPU built on SRAM-based compute-in-memory technology, further reducing application power consumption.

At the tooling level, Axon Platform comes with the ArchitStudio DSA tool, the Archit.ai architecture exploration engine, and the HiCon NoC specification tool, turning “customizable” from a concept into deliverable engineering capability. Currently, Axon Platform covers multiple scenarios including intelligent robots, industrial vision, automotive perception and smart security, satisfying both lightweight real-time inference and high-compute-intensity AI tasks.

Turning architecture design from a “craft” into “engineering”

With the topic “Axon’s ‘Architect’: ArchitStudio and Architecture Exploration Tools,” Yao Yanbin addressed how the platform is “defined.” He used a vivid analogy to explain the pain point: “Finding the disease is like taking a CT scan.” In the past, edge computing architectures faced the long-cycle contradiction between operator pain points and waterfall-style development, with architects often deciding approaches based on experience and “gut feeling.”

Scenario-Defined Architecture: FZC Tech Launches Axon Platform Fully Customized AI Computing Platform

FalconRise CTO Yao Yanbin

As the core foundation within the Axon system for empowering architects, ArchitStudio provides a “Lego-block style” construction toolset. ArchitStudio is built on the self-developed RISCAL retargetable instruction set architecture language, defining instruction binary encoding, assembly syntax, semantics and pipeline stages in one place; compilers, assemblers, simulators and debuggers are all auto-generated from it. Its compiler semantics automatically match process technology, enabling newly added instructions to be automatically scheduled by the compiler without modifying upper-layer model code, achieving “define-and-deliver.” Yao Yanbin said: “Work that previously required 3 engineers to complete in 5 months can now have its approach implemented in the first month, shortening the architecture exploration cycle from months to weeks.”

Addressing the challenges of the vast design space, strong dimensional coupling and high trial-and-error costs in edge AI architecture, Axon provides an architecture exploration platform that automatically traverses candidate architectures through a “modeling–search–PPA joint evaluation” workflow, selecting the optimal solution from the Pareto front. Taking Llama 3 8B as an example, a single architecture sensitivity analysis takes only about 17 minutes; the core cycle is shortened by about 4 times, allowing chip companies to quantitatively validate architecture decisions before tape-out and achieving a shift from “guessing architecture” to “calculating architecture.”

Computing engine: full-stack compute-in-memory technology

Lin Longyang, Assistant Professor at Southern University of Science and Technology and AI Research Scientist at FalconRise Technology, delivered a speech titled “Axon’s ‘Computing Engine’: Full-Stack Compute-in-Memory Technology.” He described compute-in-memory vividly as “shortening the distance between storage and computation,” from board level and packaging down to Macro and even Bitcell level; shortening by one order of magnitude improves energy efficiency and bandwidth by one order of magnitude.

Scenario-Defined Architecture: FZC Tech Launches Axon Platform Fully Customized AI Computing Platform

Lin Longyang, Assistant Professor at Southern University of Science and Technology and AI Research Scientist at FalconRise Technology

Lin Longyang detailed four paradigms of compute-in-memory (digital vs. analog; volatile vs. non-volatile). To address the pain points of precision and process technology, the team validated digital in-memory computing macros in both SRAM and MRAM, achieving lossless-precision computation with energy efficiency greater than 100 TOPS/W.

He emphasized: “A single compute-in-memory array alone cannot run an end-to-end model; there is a large gap between the array’s peak energy efficiency and the platform’s energy efficiency.” To this end, the Axon platform provides full-stack collaborative optimization, leveraging the AdaPIM compute-in-memory chip’s reconfigurable 4D-CIM engine and dedicated compiler to support CNN, Transformer and hybrid models, with hardware utilization reaching up to 83.12%.

Spec-definable NoC determines the upper limit of AI ASICs

With the topic “Axon’s ‘Highway’: Spec-Definable HiCon-NoC,” Lu Bin shifted the discussion from computing power to data movement. “Computing power determines the upper limit; movement determines the lower limit.” In his sharing, Lu Bin compared NoC to the “highway” of an AI SoC. Traditional shared buses suffer bandwidth collapse under multi-point concurrency, point-to-point direct connections cause wiring and area explosion, and cross-layer decoupling results in non-continuous bandwidth—under these triple dilemmas, spec-definable NoC becomes key to breaking through AI ASIC constraints.

Scenario-Defined Architecture: FZC Tech Launches Axon Platform Fully Customized AI Computing Platform

FalconRise market director Lu Bin

Hicon-NoC provides three spec-definable dimensions: routing, network and protocol interfaces. With the companion HiGUI and HiGen toolchains, the NoC specification iteration cycle is shortened from days to 2 hours, and first-pass yield exceeds 99%. The quantitative metrics are equally strong: on-chip peak bandwidth of 256GB/s and cross-core access latency below 20 nanoseconds. In a 12 TOPS edge design, Hicon-NoC interconnects the NPU array, CPU and memory with a 4x4 Mesh backbone, enabling the platform to truly run.

Sci-tech innovation empowers; finance drives progress

Shanghai Pudong Development Bank introduced its characteristic sci-tech finance model. The implementation of technological innovation cannot be separated from the precise irrigation of financial resources; the deep integration of finance and technology will drive the accelerated boom of the AI computing power industry.

Scenario-Defined Architecture: FZC Tech Launches Axon Platform Fully Customized AI Computing Platform

Luo Yipeng, Corporate Client Manager at Yitian Sub-branch of Shanghai Pudong Development Bank

When AI computing power shifts from an "arms race in training" to "large-scale deployment in inference," whoever can turn scenarios into architecture and architecture into mass production will be able to secure a position in the scramble for "architectural dominance." By using scenarios to define processor architecture, FalconRise Technology has provided its own answer with the Axon Platform, and has also filled a crucial puzzle piece for the "architecture-level customization" of the AI computing power industry—now that the platform is ready, the next step is to see the answers provided by clients and chip fabrication.

/ Read the next article /

Return to NetEase Home Download the NetEase News App

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry