A versatile RISC-V chip: integrating CPU and GPU into one core

Latest update time：2024-04-05

Reads：

????If you want to see each other often, please mark it as a star???? and add it to your collection~

Source : Content compiled from Tomshardware by Semiconductor Industry Observer (ID : icbank), thank you.

X-Silicon Inc. (XSi) has created a new RISC-V microprocessor chip architecture that combines RISC-V CPU cores with vector functions and GPU acceleration into a single chip. Jon Peddie Research reports that the CPU/GPU hybrid chips are open source and designed to handle a variety of different functions, including artificial intelligence that are typically handled by dedicated CPUs and GPUs. The problem is that it should do all this in a more efficient way.

The new hybrid CPU/GPU processor is designed to be a "do-it-all" processor. According to JPR, the industry has been seeking an open-standard GPU that is flexible and scalable enough to support a variety of markets, including virtual reality, automotive and IoT devices. This new RISC-V CPU/GPU aims to solve this problem by providing manufacturers with a single open chip design that can handle any required workload.

X-Silicon's chips differ from other architectures in that they are designed to combine the functions of the CPU and GPU into a single-core architecture. This is different from typical designs from Intel and AMD, which have separate CPU cores and GPU cores. Instead, the cores themselves are designed to handle CPU and GPU tasks. In that sense, it sounds a bit like Intel's abandoned Larabee project, which was an attempt to use x86 for graphics and other workloads.

The chip uses X-Silicon's C-GPU architecture, which incorporates GPU acceleration into the RISC-V vector CPU core. The architecture features a RISC-V vector core with a 32-bit FPU and scaler ALU. It has a thread scheduler, clipping engine, rasterizer, texture unit, neural engine, and pixel processor. The chip is designed to handle applications including artificial intelligence, high-performance computing (HPC), geometric computing, and 2D and 3D graphics.

In theory, X-Silicon's hybrid chips are able to process CPU and GPU code in the same core, which gives it a number of advantages. The chip uses the open source RISC-V ISA as the CPU and GPU, running a single instruction stream. This provides low memory footprint execution and greater efficiency since data does not need to be copied between the CPU memory space and the GPU memory space.

CPU/GPU cores can be combined together to form multi-core designs, allowing manufacturers to scale processing power as needed. In a multi-core format, multiple cores are tiled on a chip and connected using fast fabrics. Fast on-chip SRAM or eDRAM caches are also implemented in this design, which act as L2 caches that can aggregate data from multiple cores. Each core can be scheduled to run graphics, AI, video, physics, HPC or other workloads independently of other cores as needed.

With this design, X-Silicon's C-GPU architecture can run any type of CPU or GPU workload. X-Silicon claims to already have a Vulkan graphics API that works with "Fusion GPU acceleration." This will greatly aid its development and adoption on Android devices.

Because the new design is based on RISC-V, anyone can use the architecture without paying instruction set royalties - unlike x86 and ARM. If it works as expected, these chips could shake up the microprocessor industry. In theory, the standard designs currently in use are not as flexible or powerful as X-Silicon claims.

While we probably won't have to wait long to find out, it remains to be seen whether it works as well in practice as it does on paper. The software development kit will reportedly be released to early partners sometime this year.

X-Silicon’s low-power, open-standard, Vulkan-enabled C-GPU

X-Silicon Inc. (XSi) demonstrated its open-standard, low-power C-GPU architecture that combines GPU acceleration with RISC-V vector CPU cores and tightly coupled memory to form a low-power, single-processor solution. It is an open source version of its unified RISC-V vector CPU-with-GPU ISA and provides register-level hardware access through a hardware abstraction layer (HAL). The NanoTile architecture in the XSi C-GPU is designed to handle real-time data processing and dynamic graphics rendering, overcoming the limitations of traditional GPU architectures.

For more than 20 years, the industry has been seeking an open-standard GPU that is flexible and scalable enough to support a variety of markets such as AR/VR, automotive, connected IoT, and vast embedded vertical markets including robotics. X-Silicon's dynamic and highly scalable C-GPU NanoTile architecture and its ability to process multiple tasks simultaneously or sequentially is unique in this regard.

X-Silicon Inc (XSi), a San Diego-based startup founded in March 2022, launched its latest innovation: an open-standard, low-power C-GPU architecture that integrates GPU acceleration into the RISC-V vector CPU core , tightly coupled memory, providing a low-power, single-processor solution. XSi's approach brings open source to its unified RISC-V vector CPU-with-GPU ISA and provides register-level hardware access through a hardware abstraction layer (HAL). The company says this enables OEMs and content providers to customize drivers and applications with unusual customizations, unlike competitors' closed solutions. XSi believes its pioneering adoption of GPU-accelerated Vulkan on RISC-V facilitates development on Android devices, which is critical to effectively address the diverse embedded vertical market.

XSi's C-GPU uses NanoTile architecture, which the company says can meet the needs of real-time data processing and dynamic graphics rendering. Traditional GPU architectures struggle to handle dynamic data, so a novel approach is needed. XSi says its processor is designed for optimal management and rendering of dynamic content, going beyond the limitations of traditional GPUs. By integrating AI/ML computing and GPU rendering, NanoTile ensures efficiency and adaptability, making it ideal for implementing advanced graphics algorithms.

The company holds 14 patents and claims its NanoTile architecture has revolutionized AI/ML computing and GPU rendering. These patents focus on optimizing data flow between processing cores and memory, reducing latency and improving computational efficiency. Additionally, NanoTile's patented IP can be deployed in edge and cloud configurations, fostering federated models of GPU computing and providing flexibility and scalability.

XSi's open standards, low-power C-GPU architecture and NanoTile platform herald a paradigm shift in GPU technology. With support for open standards, customizable hardware access and dynamic content rendering methods, XSi believes it will set a new standard for GPU architecture, enabling developers and OEMs to unlock unprecedented levels of performance and efficiency in graphics rendering and AI/ML. Computing applications.

The RISC-V ecosystem has responded positively to the launch of a new computational graphics company that is fully committed to advancing the open standards ecosystem, the company reports.

The company plans to make its software development kit available to a select group of early development partners later this year.

Startup brings MIMD to graphics and computing

X-Silicon 由前硅谷专家组建，旨在通过能够执行 AI、HPC 和 2D/3D 图形任务的基于 RISC-V 矢量的统一图形计算引擎 (C-GPU) 彻底改变 GPU 设计。其 MIMD 架构可在单个内核内独立执行 CPU 和 GPU 代码，从而降低内存使用量并提高性能。该公司的多核布局具有快速合成器结构，可增强不同应用的数据聚合。X-Silicon 寻求通过近内存计算和新颖的硬件设置来减少 GPU 延迟，并获得 14 项专利的支持。X-Silicon面向AR/VR、嵌入式设备、汽车等领域，支持标准API和开放式编程，可实现快速开发。最初的 IP 销售针对 OEM 和超大规模厂商。

我们怎么看？一些新架构的市场已经成熟。三巨头已经从传统的 SIMD 中榨取了一切，并添加了额外的核心，例如矩阵数学（有时称为张量核心）、光线追踪干扰测试引擎、编解码器和音频处理器。X-Silicon 认为它可以通过紧密耦合的块通信 MIMD 架构来满足许多（如果不是全部）这些专用处理器需求。为此，他们将利用 RISC-V ISA。

在2022年，一群来自 AMD、高通、英特尔、ATI Technologies、戴尔和其他公司的工程师、建筑师、程序员和商业开发人员正在努力工作。开发了一种新的 GPU 设计，并意识到他们应该为自己的公司命名。作为新手，他们知道初创企业会浪费大量时间来尝试获得一个聪明、令人难忘的名称和徽标。所以选择了X。它可以是实验性的、性感的或未知的——希望这一切都是真的。

为了彻底改造 GPU 着色器核心，X-Silicon 表示正在创建一种新的可扩展的基于 RISC-V 矢量的统一计算图形引擎 (C-GPU)，该引擎可以高效地计算传统 GPU 所无法胜任的下一代工作负载类型。

此类应用包括人工智能、高性能计算、视觉、几何计算以及 2D 和 3D 图形。该公司表示，其 MIMD 架构具有独特的能力，能够在同一核心中独立运行 CPU 和 GPU 代码，提供低内存占用执行、硬件寄存器裸机编程、高性能、低功耗操作和替换等功能使用单个指令流，将传统着色器程序与适用于 CPU 和 GPU 的开源 RISC-V ISA 结合起来。

在该公司的多核设计中，多个 C-GPU 核心平铺在一个芯片上，并使用片上快速合成器结构进行连接，该结构可以将每个核心的输出动态聚合到一个公共缓冲区中，即用于图形用例的帧缓冲区或用于编解码器、视频效果处理和 AI 处理的流水线缓冲区，如下图所示。

在此设计中，快速片上 SRAM 或 eDRAM 缓存将用作二级缓存，可以聚合来自多个内核的数据。计算 RAM (C-RAM) 的常见操作在内存附近完成，将进一步减少带宽并进一步提高性能。该公司声称，每个核心都可以进行软件编程，以独立于所有其他核心来计算图形、人工智能、视频、物理、高性能计算或其他工作负载。

因此，工作负载可以并行或流水线方式实现，并在核心上同时运行，而不是在传统 GPU 上顺序运行。X-Silicon 表示，它还可以在核心上运行操作系统。

该公司声称，它还可以通过近内存计算、统一内存架构和其他新颖的硬件配置来加速计算，从而减少 GPU 固有的延迟。他们为此申请了 14 项专利。

前迪士尼/Applied Minds/Giant AI 动画行业专家 Eric Powers 评论道：“几十年来，高端动画和效果一直无法切换到 GPU 来生成最终图像。专业渲染管道软件的巨大复杂性和规模，加上专用 GPU 设计导致跨越内存和平台障碍的巨大成本，完全阻碍了我们最先进技术中 GPU 的大规模采用。集成设计使 HPC 开发人员能够就地访问直接硬件加速（例如 X-Silicon 的 C-GPU 架构），这是带领我们跨越这一边界的唯一未来。”

X-Silicon 的 C-GPU 为希望控制其计算和 GPU 命运的 OEM 提供了新兴用例的市场机会。它允许新的 API（包括自定义 API）以及为应用程序定制的生态系统，而不是引导 API 来执行它从未打算执行的任务。它不再要求原始设备制造商和制造商屈服于停滞不前的生态系统中的五巨头。该公司认为其市场机会在于新兴市场，例如具有更长电池寿命的娱乐和企业AR/VR、具有显示需求的智能嵌入式设备、需要可预测专用处理的低成本汽车显示器和模块、可穿戴设备、定制动画处理和别的。

该公司计划支持直接硬件和像素访问，因此对于低内存应用程序，不需要庞大的驱动程序。该公司表示，X-Silicon 的开放标准方法及其自己的开放编程模型将有助于快速、轻松地开发新用例以及改进现有产品。当然，X-Silicon还计划支持传统软件生态系统中的所有API，包括OpenGL ES、Vulkan、Mesa和OpenCL，但该公司还将提供硬件抽象层（HAL），允许直接访问其他人则优化开源或创建自己的驱动程序和自定义 API。这是特别有趣的，因为该架构支持新兴技术，包括传统架构不支持的新渲染模型，例如神经辐射场（NeRF）和非三角形图元。

该公司计划首先向 OEM、超大规模厂商和其他系统集成商出售 IP。首次硅片的推出日期尚未给出。

X-Silicon 将 CPU 和 GPU (C-GPU) 与单一 ISA 和开放图形操作系统 (GOS) 平台集成，可以对下一代图形渲染的整体软件开发、支持和维护产生深远影响。这可以在未来的图形领域开启一个激动人心的创新时代，为新兴和传统细分市场提供新的图形算法、性能、功耗、灵活性和成本的方法。这种方法彻底改变了图形世界，让 OEM 摆脱了传统 GPU 供应商的束缚，这些供应商提供具有复杂 API 和昂贵的传统支持的黑匣子驱动程序。

传统 GPU SIMD 架构受制于主机 CPU、操作系统和图形服务，限制了创新并有助于保持现有企业对其市场的控制。新的、新兴的、较小的垂直市场，通常得不到这些传统图形供应商的服务，可以开发和支持引人注目的图形解决方案，这些解决方案可以在产品的生命周期内升级和维护。使用新显示技术、新格式和使用范例（VR/AR、360、深度、立体、多平面全息）的下一代产品通常需要一种新的渲染方法。新的开发和部署范例还需要一致的架构，例如可从边缘扩展到云的架构。

X-Silicon 不仅仅是传统 GPU 供应商的开源替代方案，它还准备提供一种新技术图形处理框架，该框架融入了最新的人工智能和超越基于三角形的渲染，以提供自 3D 引入以来从未见过的创新平台本世纪初移动设备上的图形。

参考链接

https://www.tomshardware.com/pc-components/cpus/former-silicon-valley-vets-create-risc-v-microprocessor-that-can-run-cpu-gpu-and-npu-workloads-simultaneously

https://www.x-silicon.com/news/

https://www.jonpeddie.com/news/x-silicons-low-power-open-standard-vulkan-enabled-c-gpu/

Click here to follow and lock in more original content

END

*Disclaimer: This article is original by the author. The content of the article is the personal opinion of the author. The reprinting by Semiconductor Industry Watch is only to convey a different point of view. It does not mean that Semiconductor Industry Watch agrees or supports the view. If you have any objections, please contact Semiconductor Industry Watch.

Today is the 3726th issue shared by "Semiconductor Industry Observation" with you. Welcome to pay attention.