高通推出Adreno Neural Fusion，重大移动GPU革新
高通于9月推出基于人工智能的GPU技术Adreno Neural Fusion，旨在提升移动游戏体验。该技术是其移动GPU架构的重大革新，将AI引入图形处理，以改善性能与能效。
原文深度解读
高通于9月2日推出基于人工智能的移动GPU技术Adreno Neural Fusion，旨在提升移动游戏体验。该技术通过架构革新，将AI计算集成到GPU中，提高图形质量、帧率并降低功耗。文章涉及高通公司、其产品营销高级总监Cisco Cheng，以及Unity和Unreal Engine等游戏引擎。关键数字包括：GPU分为三个切片，每个时钟频率1.45 GHz，Matrix Cores提升性能和能效达40%，HPM内存18 MB，能效提升12%。该技术与电子元器件供应链相关，因为高通是移动芯片的主要供应商，其GPU架构变化可能影响芯片设计、制造和终端设备性能。
- 高通于9月2日推出Adreno Neural Fusion，该技术旨在通过AI提升移动游戏体验，同时降低渲染成本和功耗。
- Adreno Neural Fusion通过架构改进，将图形和AI计算集成到统一处理架构中，以消除传统移动游戏中画质、帧率和电池寿命之间的权衡。
- 高通为Adreno GPU引入新的三切片架构，每个切片以1.45 GHz时钟速度运行，并可独立执行计算。
- 每个切片包含新的Adreno Matrix Core，这是首次在GPU中引入专用AI计算核心，此前已在CPU和NPU中部署矩阵计算加速架构。
- Matrix Cores消除了图形处理期间AI计算数据在GPU外部的传输需求，与之前方案相比，性能和能效提升高达40%。
- 高通在Adreno GPU内放置了18 MB的高性能内存（HPM），允许中间计算结果在GPU内处理，减少数据移动，延迟降低，能效提升12%。
供应链影响
- Adreno Neural Fusion的架构变化可能影响高通下一代Snapdragon平台的芯片设计，进而影响相关芯片制造和封测供应链的需求。
- 新GPU架构对内存和计算资源的需求可能影响高通向内存供应商（如DRAM制造商）的采购策略，具体取决于HPM的集成方式。
- AI图形技术的集成可能提升移动设备对高性能芯片的需求，从而影响终端设备制造商（如手机品牌）的采购决策，但取决于市场接受度。
- 与Unity和Unreal Engine的集成可能促进游戏开发，进而增加对支持该技术的设备需求，可能影响相关元器件订单，但取决于开发者采用率。
- 高通的架构革新可能促使竞争对手调整GPU技术路线，从而影响整个移动芯片供应链的技术迭代和投资方向，但取决于竞争动态。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
高通于9月2日推出Adreno Neural Fusion。（图片来源：高通）
高通正在推出一种基于人工智能（AI）的图形处理单元（GPU）技术，旨在改善移动游戏体验。
高通于9月2日推出Adreno Neural Fusion。该移动GPU技术旨在提供更高的图形质量，同时降低渲染成本和功耗。
“Adreno Neural Fusion是架构全面革新的成果，旨在让计算和数据尽可能靠近硅片，”高通产品营销高级总监Cisco Cheng表示。“它旨在消除传统移动游戏中的妥协，这些妥协迫使玩家在视觉质量、流畅帧率和电池寿命之间做出选择。”
游戏中的图形质量、帧率和功耗通常涉及权衡。提高图形质量会降低帧率并增加功耗，而实现高帧率则需要降低图形质量，因为有限的计算资源必须共享。即使计算能力充足，高质量、高帧率的游戏也不可避免地会增加功耗并缩短电池寿命。
矩阵核心将性能和能效提升40%
高通为Adreno引入了新的三切片架构。（图片来源：高通）
高通通过Adreno Neural Fusion解决了这些架构限制，实现了高图形质量和高帧率，同时降低了功耗。这是通过改进GPU架构的物理结构并将图形和AI计算集成到统一处理架构中实现的。
AI技术越来越多地用于神经处理，分析游戏图像，通过AI超分辨率提升分辨率，以及生成帧。
高通为其Adreno GPU引入了新的三切片架构。GPU不再作为单个块运行，而是分为三个切片。每个切片以1.45 GHz的时钟速度运行，并可独立执行计算。
每个切片都包含一个新的Adreno矩阵核心。这些是专用的AI计算核心，标志着它们首次引入GPU。高通已在其中央处理单元（CPU）和神经处理单元（NPU）中部署了矩阵计算加速架构。
新的矩阵核心消除了在图形处理期间将AI计算数据发送到GPU外部的需要。与之前的解决方案相比，这将性能和能效提升了高达40%。此前，GPU处理3D图形渲染，而NPU单独处理AI工作负载，导致延迟和功耗浪费。
将3D渲染和AI图形技术集成到单个流水线中，还能实现更复杂的图形增强。传统的AI超分辨率会导致闪烁和重影等副作用。
闪烁是指渲染图形中边缘的闪烁或不稳定，而重影则在快速场景变化时留下残影。AI超分辨率还有一个缺点，即生成的图像看起来比原生分辨率更柔和。Adreno Neural Fusion旨在最大限度地减少这些问题，同时提供更清晰的图形。
HPM将能效提升12%
高通Adreno HPM。（来源：高通）
该架构还增加了片上内存技术，以支持高速AI计算。高通在Adreno GPU内放置了18 MB的高性能内存（HPM），允许以前必须发送到主系统内存的中间计算结果在GPU内处理。减少数据移动可降低延迟，并将能效提升12%。
HPM可用于图形子系统内的工作负载，包括基于图块的渲染（将屏幕划分为网格进行渲染）、帧缓冲（临时存储完成的图像）以及计算操作。
高通还增加了对开发者生态系统的支持。该公司已将Adreno Neural Fusion与主要游戏开发引擎Unity和Unreal Engine集成。游戏开发者无需额外开发工作或引擎更新即可应用AI图形技术。
“Adreno Neural Fusion不仅仅是添加单一功能，而是移动AI处理架构重大转变的开始，”高通表示。“矩阵加速已扩展到NPU、CPU和GPU，演变为整个平台的核心能力。”
Adreno Neural Fusion将包含在高通的下一代旗舰Snapdragon平台中。更多细节计划在9月22日至24日的Snapdragon峰会上公布。
Qualcomm introduced Adreno Neural Fusion on Sept. 2. (Photo: Qualcomm)
Qualcomm is introducing an artificial intelligence (AI)-based graphics processing unit (GPU) technology designed to improve the mobile gaming experience.
Qualcomm introduced Adreno Neural Fusion on Sept. 2. The mobile GPU technology is designed to deliver higher graphics quality while reducing rendering costs and power consumption.
"Adreno Neural Fusion is the result of an architectural overhaul designed to keep compute and data as close to the silicon as possible," Cisco Cheng, senior director of product marketing at Qualcomm, said. "It is designed to eliminate the compromises in traditional mobile gaming that forced users to choose among visual quality, smooth frame rates and battery life."
Graphics quality, frame rates and power consumption in games typically involve tradeoffs. Raising graphics quality lowers frame rates and increases power consumption, while achieving high frame rates has required reducing graphics quality because limited computing resources must be shared. Even with sufficient computing capacity, high-quality, high-frame-rate gaming inevitably increases power consumption and shortens battery life.
Matrix Cores Improve Performance, Power Efficiency by 40%
Qualcomm has introduced a new three-slice architecture for Adreno. (Photo: Qualcomm)
Qualcomm has addressed these architectural limitations with Adreno Neural Fusion, enabling high graphics quality and frame rates while reducing power consumption. This was achieved by improving the physical structure of the GPU architecture and integrating graphics and AI computing into a unified processing architecture.
AI technology is increasingly being used for neural processing that analyzes game images, upscaling to improve resolution through AI super resolution, and frame generation.
Qualcomm has introduced a new three-slice architecture for its Adreno GPU. Rather than operating as a single block, the GPU is divided into three slices. Each operates at a clock speed of 1.45 GHz and can perform computations independently.
Each slice incorporates a new Adreno Matrix Core. These are dedicated AI compute cores and mark their first introduction into the GPU. Qualcomm has already deployed matrix-compute acceleration architectures in its central processing units (CPUs) and neural processing units (NPUs).
The new Matrix Cores eliminate the need to send and receive AI compute data outside the GPU during graphics processing. This improves performance and power efficiency by as much as 40% compared with previous solutions. Previously, the GPU handled 3D graphics rendering while the NPU processed AI workloads separately, resulting in latency and wasted power.
Integrating 3D rendering and AI graphics technologies into a single pipeline also enables more sophisticated graphics enhancement. Conventional AI upscaling has caused side effects such as shimmering and ghosting.
Shimmering refers to flickering or unstable edges in rendered graphics, while ghosting leaves residual images during rapid scene changes. AI upscaling has also had the drawback of producing images that appear softer than native resolution. Adreno Neural Fusion is designed to minimize these problems while providing sharper graphics.
HPM Improves Power Efficiency by 12%
Qualcomm Adreno HPM. (Source: Qualcomm)
The architecture also adds on-chip memory technology to support high-speed AI computing. Qualcomm has placed 18 MB of High Performance Memory (HPM) inside the Adreno GPU, allowing intermediate computation results that previously had to be sent to main system memory to be processed within the GPU. Reducing data movement lowers latency and improves power efficiency by 12%.
HPM can be used within the graphics subsystem for workloads including tile-based rendering, which divides the screen into grids for rendering, frame buffers that temporarily store completed images, and compute operations.
Qualcomm has also added support for the developer ecosystem. The company has integrated Adreno Neural Fusion with major game-development engines Unity and Unreal Engine. Game developers can apply AI graphics technology without additional development work or engine updates.
"Adreno Neural Fusion is not simply the addition of a single feature, but the beginning of a major shift in mobile AI processing architecture," Qualcomm said. "Matrix acceleration has expanded across the NPU, CPU and GPU, evolving into a core capability across the entire platform."
Adreno Neural Fusion will be included in Qualcomm's next-generation flagship Snapdragon platform. Further details are scheduled to be unveiled at the Snapdragon Summit from Sept. 22 to Sept. 24.