擒数网 （随信APP) | 高通发布了骁龙 8 至尊版，主频超过 4GHz，Benchmark 多核得分超过 1 万。
文章目录[隐藏]
擒数网 （随信APP) | 高通发布了骁龙 8 至尊版，主频超过 4GHz，Benchmark 多核得分超过 1 万。
【微信/公众号/视频号/抖音/小红书/快手/bilibili/微博/知乎/今日头条同步报道】
高通的骁龙旗舰芯片这几年的命名几经变更，比如前几年突然出现的骁龙 888，主要图一个吉利，后面变成了骁龙 8 Gen1、Gen2、Gen3，到了今年骁龙移动芯片的命名开始和 PC 端命名接近，英文名 Snapdragon 8 Elite，中文名骁龙 8 至尊版。
取这个名字不光要勇气，其实也需要一点底气，不过从这一代骁龙 8 系旗舰芯片的账面表现来说，说是至尊问题也不大。
### Oryon CPU 架构初来移动端，全大核，跑分强
以往的移动端芯片设计许多用到了 big.LITTLE 架构，即大小核架构，由此还可以派生出「超大小」架构，用超大核和大核（性能核心）负责重载任务，比如大型游戏，用小核（能效核心）负责轻型任务，比如聊天和网页浏览等。
不过我们也可以看到，移动端芯片的小核数量一直减少，骁龙 8 Gen2 有 3 个小核，到了骁龙 8 Gen3 变成了 2 个小核，今年的骁龙 8 至尊版干脆直接取消小核，变成了全大核架构，2 个超大核心配合 6 个大核心。
去年高铁把自研的 Oryon CPU 架构带到了骁龙 X Elite 上，今年第二代高通 Oryon CPU 架构来到了骁龙 8 至尊版上，这也是这次这款旗舰芯片架构变化，以及性能暴增的主要原因。
首先就是骁龙 8 至尊版超大核的主频达到了惊人的 4.32GHz，这样的主频出现在移动端芯片上确实让人意外，毕竟这样的数字一般是出现在桌面芯片上。
哪怕是另外的 6 个大核，其主频也达到了 3.53GHz。
在这样的架构加持下，骁龙 8 至尊版的试验机型在 Benchmark 6 的跑分中能够获得单核超 3200，多核超 10000 的上佳成绩。
作为对比，上代骁龙 8 Gen3 在 Benchmark 官方网站上的典型成绩是单核 2213，多核 7466。
上上代骁龙 8 Gen 2 的 Benchmark 6 典型成绩则为单核 2036，多核 5697。
所以说这次骁龙 8 至尊版的 CPU 性能是挤爆牙膏一点儿问题也没有，按照高通的说法，骁龙 8 至尊版的单核性能和多核性能相较于上代都提升了 45%，同时能耗也降低了 40% 以上。
另外为了适应 AI 时代的需要，这代 Oryon CPU 的内存架构也有巨大的升级，每个 CPU 丛集都有 12MB 的 L2 缓存，另外超大核和大核的 L1 缓存也比之前有所增加，同时骁龙 8 至尊版也支持目前最先进的 LPDDR5x 内存。
### 捆绑虚幻 5 引擎，GPU 也变强了
我们知道游戏性能表现是 CPU 和 GPU 一起合力的结果，所以这一次 CPU 的性能提升自然也会提升游戏性能表现，高通举的例子是在虚幻引擎的 Chaos Physics 系统，这是一个旨在赋能下一代游戏的模拟工具，在这个工具中可以在终端侧实时模拟所有必要的游戏物理系统。因为这代骁龙 8 至尊版的 CPU 性能加强，它支持在这个模拟工具中超过 1000 个物体碰撞和破坏的场景，并且时延低于 5 毫秒。
和 CPU 架构更新类似，骁龙 8 至尊版也首次引入了 Adreno GPU 切片架构，让 GPU 性能提升了 40%，功耗也降低了 40%。
高通也演示了运行虚幻 5 引擎的 Nanite 方案的视频，目前仅有骁龙平台支持这个功能，借助这个功能，理论上游戏开发者可以在移动游戏里渲染出电影级别的 3D 环境。
在过往，我们看到了骁龙旗舰芯片在游戏上的诸多特性，一边是为了降低能耗保证基础画质体验的特性，比如可变分辨率，和智能插帧等等，另一边是为了提高画质上线，比如移动端的光线追踪等等。
这一次高通和虚幻 5 引擎在发布会上的强绑定，似乎透露出让移动游戏体验逼近主机级或者 PC 级游戏的野心，而在发布会 Keynote 上出现的《永劫无间手游版》和《剑侠情缘网络版三无界版》等游戏，也倾向于让移动游戏和 PC 平台的玩法与数据接近，甚至互通。
### AI 与影像，逐渐融合，逐渐强大
去年骁龙 8 Gen3 发布的时候，AI 部分的演示集中在了端侧大语言模型的文生文，还有 Stable Diffusion 文生图，源头都是文字。
到了今年，因为 AI Agent 概念的火爆，所以 AI 应用也呈现出了非常明显的多模态化趋势，而不再是文生一切。
因为 AI 任务非常多样，所以骁龙 8 至尊版的 AI 引擎也得到了升级，Oryon CPU 主要负责时延敏感型的任务，以及进行 AI 任务的分配，比如把任务分配到传感器中枢去。
Hexagon NPU 是高通 AI 引擎的中心，高通给 NPU 里的标量和向量加速器增加了额外的内核，现在向量加速器有 6 个核心，标量加速器有 8 个核心，张量处理器也得到了加强，这让 NPU 的整体推理速度提升了 45%，面对 AI 多任务需求，高通也增强了并发功能，可以让 AI 和计算机视觉负载可以在 NPU 内存中共存。
CPU、GPU 和 NPU 都得到了巨大的性能提升之后，那么高通的目的也非常明显，那就是在终端侧能够运行多模态 AI 助手，运行多个 AI 模型，包括 LMM（多模态大模型）、LLM（大语言模型）、LVM（视觉大模型）、ASR（自动语音识别）等等。
举个例子，之前用户使用语音助手，总感觉语音助手反应慢半拍，但是现在的旗舰手机语音助手的交互更灵敏更像真人，几乎没有延迟。这背后就是技术的进步。
之前的语音助手需要把语音转换成文字，让 AI 模型理解文字，整理成文本，然后大语言模型处理文本，再把回答生成文本，文本再转换成语音输出给用户。
现在的一些 AI 助手可以直接理解语音，无需把语音转换成文本，大大缩短了处理时间。
除了对文本和语音的理解能力加强，现在在骁龙 8 至尊版的加持下，AI 助手也可以在终端侧直接用摄像头理解看到的内容，比如用相机对准收据，AI 助手就能理解收据内容，推理出信息结果，帮助用户计算付款信息，如何和朋友分担账单等等。
除了摄像头和麦克风这样明显的传感器之外，其他的传感器也能实时了解周遭环境，综合起来根据用户的具体喜好和情况来创建用户个人神经网络，从而实现更贴合用户习惯，更洞察用户潜在需求的效果。
骁龙 8 至尊版的 ISP（图形信息处理器）得到了增强：全新的 AI ISP 和在底层和 NPU 协作，大大提升像素吞吐量，达到了每秒 43 亿像素，骁龙 8 至尊版支持 3 个 4800 万像素的相机同时进行 30fps 的视频拍摄。
另外，因为 Hexagon NPU 可以和内存直连，这个直连技术也让 NPU 能够直接访问 ISP 上的原始 RAW 数据，从厄尔实现对 4K 60fps 拍摄中的实时 AI 辅助增强，实现各种 AI 效果。
将 AI 算法直接引入图像和视频处理管线之后能够获得各种各样的 AI 功能和效果，比如无限语义分割（Limitless Segmentation），支持对图片进行超过 250 层语义识别和分割，对照片的所有部分进行专门的调整和改善。
一些实际的 AI 视频效果包括 AI 补光，哪怕是在视频会议中处于严重背光环境，用户也可以添加虚拟光源，保证脸部画面清晰。另外一个刚需功能就是 AI 视频魔法消除，现在已经有不少手机支持 AI 照片消除，把照片里的不必要杂乱元素擦除，而在 AI 视频魔法消除中，用户可以在视频里选择需要擦除的对象，从而在整个视频里保证画面的干净，而这一切都在骁龙 8 至尊版终端侧实现，无需云端计算。
2024 年手机行业出现了复苏迹象，AI 手机概念也有可能迎来转折点，去年我们看到的更多是不太可用的文生文和文生图演示，而今年我们则看到了更多具体的，能用到并且有用的服务，这一切又刚好迎来了高通骁龙 8 至尊
英文版:
Over the past few years, Qualcomm's Snapdragon flagship chips have undergone several name changes. For example, a few years ago, the suddenly appeared Snapdragon 888 was mainly for good luck. Later, it became Snapdragon 8 Gen1, Gen2, Gen3. This year, the naming of Snapdragon mobile chips started to resemble that of PC chips, with the English name Snapdragon 8 Elite and the Chinese name Snapdragon 8 Ultimate Edition.
Oryon CPU Architecture Debuts on Mobile, All Big Cores, Strong Performance
In the past, many mobile chip designs used a big.LITTLE architecture, which includes big and small cores. From this, it can also derive the "super big" architecture, where super big cores and big cores (performance cores) handle heavy tasks, such as large games, and small cores (efficiency cores) handle light tasks, such as chatting and web browsing.
However, it can also be seen that the number of small cores in mobile chips has been decreasing. Snapdragon 8 Gen2 had 3 small cores, which decreased to 2 small cores in Snapdragon 8 Gen3. This year's Snapdragon 8 Ultimate Edition directly eliminated the small cores and became an all big core architecture, with 2 super big cores and 6 big cores.
Last year, Qualcomm introduced its self-developed Oryon CPU architecture on the Snapdragon X Elite. This year, the second-generation Qualcomm Oryon CPU architecture has arrived on the Snapdragon 8 Ultimate Edition. This is the main reason for the architectural changes and the significant performance increase of this flagship chip.
First of all, the main frequency of the super big core in the Snapdragon 8 Ultimate Edition has reached an astonishing 4.32GHz. Such a main frequency on a mobile chip is indeed surprising, as such numbers are usually seen on desktop chips.
Even the other 6 big cores have a main frequency of 3.53GHz.
With this architecture, the test model of the Snapdragon 8 Ultimate Edition can achieve single-core scores exceeding 3200 and multi-core scores exceeding 10000 in Benchmark 6.
For comparison, the typical scores of the previous Snapdragon 8 Gen3 on the Benchmark official website were 2213 for single-core and 7466 for multi-core.
The typical scores of the previous Snapdragon 8 Gen 2 on Benchmark 6 were 2036 for single-core and 5697 for multi-core.
Therefore, the CPU performance of the Snapdragon 8 Ultimate Edition this time has no problem with squeezing toothpaste. According to Qualcomm, the single-core performance and multi-core performance of the Snapdragon 8 Ultimate Edition have increased by 45% compared to the previous generation, while power consumption has also decreased by over 40%.
Paired with Unreal Engine 5, GPU Also Strengthens
We know that gaming performance is the result of the combined efforts of CPU and GPU. So this time, the performance improvement of the CPU will naturally enhance gaming performance. An example given by Qualcomm is the Chaos Physics system in the Unreal Engine, which is a simulation tool designed to empower next-generation games. In this tool, more than 1000 object collisions and destruction scenes can be simulated in real-time on the terminal side, with a latency of less than 5 milliseconds, supported by the enhanced CPU performance of the Snapdragon 8 Ultimate Edition.
Similar to the CPU architecture update, the Snapdragon 8 Ultimate Edition also introduces the Adreno GPU slice architecture for the first time, increasing GPU performance by 40% and reducing power consumption by 40%.
Qualcomm also demonstrated a video running on the Unreal Engine's Nanite solution. Currently, only the Snapdragon platform supports this feature. With this feature, theoretically, game developers can render movie-level 3D environments in mobile games.
In the past, we have seen many features of Snapdragon flagship chips in games, including features for reducing power consumption to ensure basic image quality, such as variable resolutions and intelligent frame interpolation, as well as features for raising the image quality bar, such as mobile light chasing, etc.
This time, the strong integration between Qualcomm and Unreal Engine 5 at the conference seems to reveal an ambition to bring mobile gaming experience closer to the level of console or PC games. In addition, games like "Endless Eternal Mobile Edition" and "Swordsman Love Online Version 3D" that appeared at the keynote seem to aim to bring mobile games closer to PC platform gameplay and data, even to interchange.
AI and Imaging, Gradually Integrating, Gradually Strengthening
When the Snapdragon 8 Gen3 was released last year, the AI demonstrations focused on end-to-end large language models and image-to-image stable diffusion systems, all of which revolved around text.
By this year, due to the popularity of AI Agent concepts, AI applications have shown a very clear trend towards multimodalism, rather than just focusing on text.
Because AI tasks are very diverse, the AI engine in the Snapdragon 8 Ultimate Edition has also been upgraded. The Oryon CPU is mainly responsible for latency-sensitive tasks and AI task allocation, such as distributing tasks to sensor hubs.
The Hexagon NPU is the central hub of Qualcomm's AI engine, with additional cores added to the scalar and vector accelerators inside the NPU. The vector accelerator now has 6 cores, the scalar accelerator has 8 cores, and the tensor processor has also been strengthened, increasing the overall inference speed of the NPU by 45%. In the face of AI multitasking demands, Qualcomm has also enhanced concurrent functionality, allowing AI and computer vision workloads to coexist in the NPU memory.
With significant performance improvements in the CPU, GPU, and NPU, Qualcomm's goal is quite obvious - to run multimodal AI assistants at the terminal side, running multiple AI models, including LMM (multimodal large model), LLM (large language model), LVM (visual large model), ASR (automatic speech recognition), and more.
For example, in advance, users used to feel that voice assistants were slow to respond. But now, flagship phone voice assistants are more responsive and more human-like, with almost no delay. This progress is due to technological advancements.
Previously, voice assistants had to convert speech to text, let the AI model understand the text, organize it into text, then have the large language model process the text, generate the response, convert the text back to speech, and output it to the user. Now, some AI assistants can directly understand speech without converting it to text, greatly reducing processing time.
In addition to strengthening the understanding capabilities of text and speech, with the support of the Snapdragon 8 Ultimate Edition, AI assistants can now directly use the camera at the terminal side to understand what they see. For example, aiming the camera at a receipt, the AI assistant can understand the receipt's content, infer the information, help the user calculate payment information, and how to split the bill with friends, among other tasks.
In addition to obvious sensors like camera and microphone, other sensors can also understand the surrounding environment in real-time, creating a user-specific neural network based on the user's specific preferences and circumstances, leading to a more personalized user experience and insights into user needs.
The ISP (Image Signal Processor) of the Snapdragon 8 Ultimate Edition has been enhanced: the new AI ISP works in cooperation with the NPU at the lower level, significantly increasing pixel throughput to 4.3 billion pixels per second. The Snapdragon 8 Ultimate Edition supports simultaneous video shooting at 30fps with 3 48-megapixel cameras.
Moreover, the direct connection of the Hexagon NPU to memory allows the NPU to directly access the raw RAW data on the ISP, enabling real-time AI enhancement in 4K 60fps shooting, achieving various AI effects.
After directly integrating AI algorithms into the image and video processing pipeline, various AI functions and effects can be obtained, such as Limitless Segmentation, which supports identifying and segmenting over 250 layers of semantics in images and making specific adjustments and enhancements to all parts of a photo.
Some practical AI video effects include AI Fill Light, where even in heavily backlit environments, users can add virtual light sources to ensure clear facial images in video conferences. Another important feature is AI Video Magic Erasure, where many smartphones now support AI photo erasure, removing unnecessary clutter from photos. In AI Video Magic Erasure, users can select objects to erase in the video, ensuring a clean picture throughout the video. And all of this is achieved at the terminal end of the Snapdragon 8 Ultimate Edition, without the need for cloud computing.
In 2024, the mobile industry showed signs of recovery, and the concept of AI phones may also reach a turning point. Last year, we saw more demonstrations of text and image that were not very useful, while this year we see more specific and useful services that can be applied and have utility. All of this coincides with the significant performance improvement of the Snapdragon 8 Ultimate Edition, the collective design shift towards the Android camp, and the earlier overall release schedule.
The result of fierce competition is that product improvements cannot be squeezed out, and users have more choices.
Author: Liu Xuewen
#Follow the official WeChat account of iFanr: iFanr (WeChat ID: ifanr), offering you more exciting content in real-time.
iFanr | Original Article · View Comments · Sina Weibo
高通骁龙 8 至尊版发布，主频超 4GHz，Benchmark 6 多核破 1 万
#高通骁龙 #至尊版发布主频超 #4GHzBenchmark #多核破 #万
关注流程：打开随信App→搜索擒数网随信号：973641 →订阅即可！
公众号：擒数网 抖音：擒数网
视频号：擒数网 快手：擒数网
小红书：擒数网 随信：擒数网
百家号：擒数网 B站：擒数网
知乎：擒数网 微博：擒数网
UC头条:擒数网 搜狐号：擒数网
趣头条：擒数网 虎嗅：擒数网
腾讯新闻：擒数网 网易号：擒数网
36氪：擒数网 钛媒体：擒数网
今日头条：擒数网 西瓜视频：擒数网
