Away from the performance headliners and press articles about the cloud industry, there is a vast opportunity for tinyML at the far edge.  This is a diverse collection of sensor-rich endpoints that collect, pre-process, and analyze real-world data prior to submitting metadata to the cloud. Learn about how SiFIve products are driving innovation across this space.


**SiFive’s X100-Series, Part of the Second Generation SiFive Intelligence Family**

Right now, there is a massive focus on “up and to the right” hardware technologies that can push AI performance levels to new heights for data centers. During the HotChips 2025 conference, the thirst for improved performance in hyperscalers seemed to drive the vast majority of the agenda, as the big, powerful chips stole the spotlight.

Yet, away from the performance headliners and press articles about the cloud industry, there is a vast opportunity for tinyML at the far edge.  This is a diverse collection of sensor-rich endpoints that collect, pre-process, and analyze real-world data prior to submitting metadata to the cloud.  This graph forecasts the fast-growing AI workloads that will be run in edge, mobile, PC, and IOT systems markets. 

![graph](https://images.prismic.io/sifive/aMM5K2GNHVfTPEW__graphforblog.png?auto=format,compress)

The reason behind this growth is simple: not all AI tasks can or should be done in the cloud or a large datacenter.  Massive opportunities exist where there is inadequate internet access, need for real-time responsiveness, and a need for privacy that prevents cloud-based solutions that have network dependencies, data latencies, and security/privacy concerns. Even if cloud-based solutions are used, the ability to do AI locally reduces data movement costs and power consumption and is more sustainable. While other IP suppliers have “paused” their product developments for this set of use cases, SiFive is doubling down and investing heavily in broadening its offering for the diverse set of use cases at the far edge and complementing those processors with reference AI software stacks. The focus here is on enabling our customers to create and deploy leadership products for their areas of focus in the most cost- and time-effective way.

![AI models effectiveness](https://images.prismic.io/sifive/aMM5LGGNHVfTPEXA_slideforblog.png?auto=format,compress)
 
**What does it take to bring efficient AI to the edge?**
 
With the pace of innovation of AI models (with convolutional neural networks being replaced with transformer models, and some industry experts already predicting the death of those), building a flexible architecture that can adapt is key.  To enable AI everywhere, the companies developing products for edge use cases are focused on improving latency, optimizing usage of bandwidth, and enhancing data security. These systems are typically faced with intensive resource constraints, including power, cost, and size.  Many vertical markets have additional unique needs: in healthcare, complying with data privacy laws is paramount; in industrial and manufacturing sites, real-time responsiveness is crucial; in automotive, connectivity is not guaranteed.

The industry uses a wide range of names for this broad set of applications, with one of the more recent ones being “Physical AI”. Whatever market categorization is used, at SiFive, we embrace Local AI and are showing why harnessing open-standard RISC-V as the foundation of these systems is the best solution for these applications - with advantages in power, size and common software. Starting with just one product in 2021, the first generation of our Intelligence product family has become very popular, with over 40 design wins across a diverse set of IoT applications. To meet increasing and broader market demand, the 2nd generation Intelligence products launched officially today.  The combination of dedicated scalar and vector pipelines in our new Intelligence products provides our customers choice, scalability and an opportunity to innovate in the creation of accelerated compute solutions. The smallest of these, the [new X100-series](https://www.sifive.com/blog/the-perfect-solution-for-local-ai) were designed based on direct feedback from customers about the need for more efficient AI inference at the deeply embedded edge.

![AI products chart](https://images.prismic.io/sifive/aMM5KmGNHVfTPEW-_blog3.png?auto=format,compress)

Our products do not just operate as standalone processors but can be paired with customer accelerators.  While specialized accelerators can deliver great performance and efficiency, the accelerator function is fixed and can be limited. As SiFive co-founder Krste, mentions in his [product launch video](https://www.youtube.com/watch?v=_oTL_8IV5Ho), one of our key advantages with the Second Generation Intelligence family, is to enable the creation of ‘intelligent accelerators’, combining powerful fixed function accelerators with a programmable front end. This is accomplished by incorporating two direct-to-core interfaces specifically designed for data movement and control functions. 

Another aspect of AI is the datatype support. With all second gen Intelligence products, we now have support for Bfloat16; originally developed by Google Brain, Bfloat16 was designed specifically for the needs of neural network training and inference. Importantly, this data type allows AI inference to achieve a level of accuracy comparable to the standard 32-bit format while reducing memory usage and speeding up computations. This is an example of a function that has only recently started being deployed at scale and, therefore, isn’t supported on many legacy products. While I focus here on the capabilities in the Second Generation of the Intelligence Family, you can imagine that SiFive is actively tracking other industry standards and proposals about data types and engaging with customers to determine the future direction of this product family.

There are two members of the X100 product series at this point, the [X160](https://www.sifive.com/document-file/sifive-intelligence-x160-gen2-product-brief) and [X180](https://www.sifive.com/document-file/sifive-intelligence-x180-gen2-product-brief). The latter is a 64-bit processor. One piece of early customer feedback was that it would typlically be paired with a 64-bit host processor (typically running Linux). Use of the X180 simplifies the memory map and sharing of information between the cores.  

So, what does all this mean for vertical markets that are developing and adopting Local AI solutions?  Dedicated vector processing in small form factors means AI workloads can be run faster, and with less power, ultimately enabling more diverse solutions that deliver real-time responses with enhanced privacy, without requiring network connectivity.  Having the option of a single scalable architecture that can span the breadth of edge AI applications makes product development and software programming much simpler.  Product roadmaps can be future-proofed for the rapidly changing AI model landscape as customized instructions can be added.  Instead of being restricted to limited processor options and siloed software stacks, the RISC-V open standard means your product engineers can think big about tinyML.

The opportunity is so great, and the advantages so compelling that, even before being launched, the products have already been licensed by two top tier US semiconductor manufacturers to build into their future edge AI SoCs.   You can read more on these case studies [here](https://www.sifive.com/document-file/x160-standalone-vector-cpu-case-study), and you can read more details about the [X100 product series](https://www.sifive.com/cores/intelligence-x100-series).  

If you want to discuss how SiFive products can be part of your company’s AI journey, please [contact us](https://www.sifive.com/contact-sales). 

Ian Ferguson, VP Vertical Markets at SiFive

Blog describes AI products from SiFIve driving innovation at the far edge of AI compute

SiFiive-Intelligence-Gen2-Banner.png

近年来，AI 一直是科技行业的焦点。随着 RISC-V 的快速发展，SiFive 通过我们的 Intelligence IP 家族取得了领先地位，提供了一个基于单一 ISA 的可扩展计算平台，并具备根据特定 AI 工作负载进行定制的能力。

**全新 X100 系列，SiFive Intelligence 系列第二代成员**

近年来，AI 一直是科技行业的焦点。随着 RISC-V 的快速发展，SiFive 通过我们的 Intelligence IP 家族取得了领先地位，提供了一个基于单一 ISA 的可扩展计算平台，并具备根据特定 AI 工作负载进行定制的能力。

我们经常收到客户的需求，就是希望能把 AI 能力从云端移到工厂、家庭、办公室，乃至人们手中——让 AI 本地化。他们之所以有这样的需求，主要有几个原因。首先，在本地处理数据可以实现更快的响应，使人与设备的交互更加自然。其次，这也有助于保障数据的隐私和安全——如果数据从未离开本地设备，就更容易掌握谁在访问以及他们如何使用数据。随着我们越来越依赖这项技术，确保它在网络连接或基础设施不可用时依然能稳定运行，也变得尤为重要。
然而，要实现这一点并非易事。AI 通常需要大量数据和高强度的计算能力，而现有的嵌入式处理器往往难以满足所需的数据处理性能；另一方面，受成本与功耗限制，这些设备又无法采用专为数据中心设计的高性能 CPU。SiFive 的全新 X100 系列通过几个关键方面可以解决这些挑战。

首先，我们将应用在 X200 和 X300 系列上的向量引擎进行了缩减，使其能够满足严格的成本和功耗限制，同时仍然提供足以支持众多边缘 AI 推理场景的数据处理性能。RISC-V 向量 ISA 以及一系列巧妙的实现特性，造就了灵活的向量引擎，即便在低成本、高延迟的存储环境下，也能很好地应对这些挑战。
这些特性结合在一起，使得基于 SiFive X160 的设计在执行许多 AI 推理任务时，速度超过 Arm™ Cortex®-M85 的两倍，而硅片面积却是相同的。

![SiFive vs. Arm Chart](https://images.prismic.io/sifive/aL2oqGGNHVfTOvXU_figure1.png?auto=format,compress)
**图 1：SiFive Intelligence X160 在典型 AI 任务中的性能对比** 

不过，也有一些边缘 AI 任务需要更高性能，而要在边缘设备的成本和功耗预算内实现这一点，就必须引入定制加速器。这样客户就能在最小的面积和功耗约束下，打造出正好符合目标任务需求的计算配置。然而，将这些加速器集成到复杂的计算子系统中，在硬件和软件层面都可能面临困难和效率问题。为了简化这一过程并提升效率，我们提供了两种不同的与内核直连的协处理器接口，可以单独或结合使用，取决于最终应用的需求。这让 X100 系列能够作为加速器控制单元 (ACU) 来使用，承担起与客户定制加速器进行细致协作和管理的角色。

虽然 X100 系列才刚推出，但已有不少客户积极将其集成到系统中，以利用其无与伦比的 AI 效率。其中包括两家美国大型半导体公司：一家利用 X160 的向量性能驱动下一代边缘 AI SoC，另一家则使用 X160 来控制并协助其自研的 AI 加速器。

除了 X160，我们还发布了 X180。这款 Intelligence IP 实现了 64 位 RV64I ISA，具备更大的物理地址空间，非常适合与更大型的 CPU 一起在复杂 SoC 中协作。

![SiFive Intelligence 160 diagram](https://images.prismic.io/sifive/aL2op2GNHVfTOvXS_fig2.png?auto=format,compress)
![SiFive Intelligence 180 diagram](https://images.prismic.io/sifive/aL2op2GNHVfTOvXT_fig21.png?auto=format,compress)       
**图 2：SiFive Intelligence X100 系列产品：X160 与 X180**

这些新产品非常适合支持本地 AI、边缘推理，以及深度嵌入式远端场景下的向量处理。若想进一步了解这些产品、查看规格，或阅读客户/行业应用案例，[请访问 Intelligence 网页](https://www.sifive.cn/cores/intelligence)，并下载 [X160](https://www.sifive.cn/document-file/sifive-intelligence-x160-gen2-product-brief)、[X180](https://www.sifive.cn/document-file/sifive-intelligence-x180-gen2-product-brief) 以及 [Intelligence 系列的产品简介](https://www.sifive.cn/document-file/sifive-intelligence-gen2-family-brief)。

如果您正在研发本地 AI 解决方案，并想探索这些或其他 SiFive 产品是否适合您的项目，请联系 [SiFive 销售团队](https://www.sifive.cn/contact-sales)。

本地 AI 的完美解决方案

SiFive_intelligence-gen2_Blog_1000x500-v1.jpg

RISC-V 开发者支持与生态建设更新：Red Hat RHEL 已支持 HiFive Premier P550、NVIDIA CUDA 移植、RVA23 正式通过、Canonical Ubuntu 支持以及 RISE 项目的进展。

![Premier P550 Board](https://images.prismic.io/sifive/aJqr6aTt2nPbaMH9_SiFive_NVIDIA-CUDA_Blog_1000x500.jpg?auto=format,compress)
**作者：Ian Ferguson，SiFive 商务拓展副总裁** 

我在 2024 年初加入 SiFive 时，最初的工作重点（此后有扩大）是研究如何推动 SiFive 的 [RISC-V 技术](https://www.sifive.cn/about/risc-v-history)在面向[数据中心](https://www.sifive.cn/solutions/data-center)的 system-on-chip 组件中可以被更广泛采用。引入新技术并取代长期占据主导地位的既有厂商，这并非几个季度就能完成的任务，这类计划往往需要多年才能真正落地。我深有体会：早在 2008 年初，我和几位同事（我们在公司内部被称为“特立独行者”，但我们更喜欢“有远见者”这个称呼）共同发起推动 Arm 进入服务器市场的计划。

加入 SiFive 后，我很快意识到，我们的 CPU roadmap 在单线程性能方面终将赶上现有供应商的产品。但要真正向更广泛的客户群开放这个市场，而不仅仅是一两个早期[创新者](https://medium.com/@samuelvandeth/crossing-the-chasm-162802d1cf27)，还需解决另外三个方面的问题。

在这篇文章中，我将先探讨第一个方面：如何赋能软件开发者社区。至于另外两个方面，我暂时先保留起来，在我后续的文章中会再讨论它们。

**三个数字**

**90**

**4**

**12,800**


这些数字代表什么？目前，预测显示 NVIDIA 占有 90% 的基于 GPU 的神经网络训练市场。他们拥有超过 4 百万名的开发者。而在 Linux 内核中，每天平均有 12,800 行代码被新增、修改或删除。

有人说“硬件再次变得酷炫”，但归根结底，若缺乏强大的软件支持，企业要采用新硬件技术的成本和难度仍然过高。


今年 5 月，[Red Hat 宣布](https://www.redhat.com/en/blog/red-hat-partners-with-sifive-for-risc-v-developer-preview-for-red-hat-enterprise-linux-10)发布 Red Hat Enterprise Linux (RHEL) 开发者版本，用于 SiFive 最新的 [RISC-V 开发板 HiFive Premier P550](https://www.sifive.cn/boards/hifive-premier-p550)。此外，他们还向 [CentOS 社区](https://blog.centos.org/2025/05/initial-centos-support-for-risc-v/)开放了 RISC-V 源代码。

7 月，NVIDIA 宣布正在将 CUDA 移植到 RISC-V，并以 HiFive Premier P550 作为起点。尽管他们尚未公布代码可用的时间表，但该消息已在媒体、客户以及竞争对手之间引发广泛关注。

我只能分享目前已公开的信息。Joe Byrne 的[一篇精彩分析](https://xpu.pub/2025/07/23/cuda-risc-v/)則提供了有用的背景信息。正如 Joe 写道：

“NVIDIA 正将 RISC-V 与 Arm 和 x86 一起作为受支持的架构——这意味着，一旦代码可用，企业将能够选择 RISC-V CPU 与 NVIDIA GPGPU 结合使用。

RISC-V International 也刚发布一篇 [blog](https://riscv.org/blog/2025/08/nvidia-cuda-rva23/)，內容基于他们与 NVIDIA 的 Frans Sijstermans 的对话，分享这则在 [RISC-V 中国峰会](https://riscv-summit.com/en/)上公布的重要消息。


**来自 Red Hat 和 NVIDIA 的共同主题**

我认为 Red Hat 与 NVIDIA 的公告有三个共通点：

**面向未来：** 两家公司都在为 RISC-V 服务器硬件广泛部署的未来做准备。他们希望能提前做好准备。

**首选平台：** 两家公司都以 [HiFive Premier P550](https://www.sifive.cn/boards/hifive-premier-p550) 开展初步工作。这归功于 HiFive Premier P550 相较于其他 RISC-V 开发板更高的性能、稳定性，以及 SiFive 工程团队在上游开源代码方面的出色工作，此外还有硬件本身的性能和扩展能力。这种强大的芯片功能使许多开发人员能够以全新方式 “亲身体验” RISC-V。

**RVA23 里程碑：** [RVA23 配置文件](https://riscv.org/riscv-news/2024/10/risc-v-announces-ratification-of-the-rva23-profile-standard/)的正式通过产生了极大的推动作用。大型生态系统参与者希望将目标锁定在由多个供应商支持、且具备一致处理器架构的平台上。而软件可依赖的通用指令集是实现这一目标的关键。

同时，也必须认可 Canonical 在 RISC-V 生态中的开创性贡献，他们让 Ubuntu 成为 RISC-V 的一等公民。

这些行业巨头的共同投入，加上众多参与 [RISE 项目](https://riseproject.dev/)的公司支持，都在推动一个现实：客户将能够从三种处理器架构中自由选择。系统构建商、软件开发者和组件供应商面临的障碍几乎每天都在被逐步消除。

RISC-V 的独特在于它的开放性。供应商可以针对狭窄且高度专业化的垂直市场开发芯片。NVIDIA 显然看到了将 CUDA 扩展到全新应用领域的机会——这对于其他商业化、为广泛通用市场而设计的指令集架构（ISA）是很难实现的。CUDA 的强大功能与 RISC-V 的无限未来相结合确实令人兴奋。

SiFive 最近刚庆祝[成立十周年](https://www.sifive.cn/blog/sifive-celebrates-10-years-as-your-trusted-partner)，我们将继续突破 RISC-V 的极限。一如既往，我们想听到您的想法、您的创新以及您正面临的挑战。

今年 10 月，SiFive 也将参与在圣克拉拉举办的 RISC-V 峰会，期待许多精彩的演说与讨论。让我们利用这个活动来扩大对话，并进一步加速 RISC-V 发展。

最后，能否证明 NVIDIA 的选择是正确的，就取决于整个 RISC-V 社区了。

RISC-V 基础设施：现在，关键在开发者

SiFive_NVIDIA-CUDA_Blog_1000x500.jpg

SiFive Blog 是您获取有关处理器IP、计算密度、芯片架构和创新突破最新消息的首选来源，无论是制造出色的可穿戴设备和消费电子产品，优化数据中心，还是构建下一代汽车。请经常查看，第一时间了解我们发布的最新内容。

获取来自领导者和RISC-V创始人的最新洞见

SiFive Blog 是您获取有关处理器IP、计算密度、芯片架构和创新突破最新消息的首选来源，无论是制造出色的可穿戴设备和消费电子产品，优化数据中心，还是构建下一代汽车。

SiFive 博客 - 来自 RISC-V 专家的最新洞察与深度技术解析

来自 RISC-V 专家的最新洞察与深度技术解析

SiFive 博客是您获取 RISC-V 最新动态的首选平台，内容涵盖处理器 IP、芯片架构、软件开发及各类技术创新。无论您正在打造下一代消费电子产品、优化数据中心，还是开发新一代汽车，都欢迎关注並获取专家们的最新见解。

SiFive Data Center Innovation

Leadership in Embedded Markets

Essential 系列

Intelligence 系列

Performance 系列

Automotive 系列

现已上市

旧版开发板

All Boards

Featured Blog Post

HiFive P550 Review

Software Expansion

Premier P550 Boards

RISC-V Software Expansion

Premier P550 Boards

Cycuity Partnership

SiFive’s New Look

SiFive 博客

Enabling AI Innovation at The Far Edge

本地 AI 的完美解决方案

RISC-V 基础设施：现在，关键在开发者