具有相同内核的 ARM SoC 之间有很大差异吗？
具有相同内核的两个 ARM SoC 之间（在性能方面）有很大差异吗？
例如，以下之间可能存在哪些差异（除了可归因于不同时钟速率的差异）：
- Apple A5，基于 1GHz 双核 ARM Cortex-A9
- 三星 Exynos GT-I9100，基于 1.2 GHz 双核 ARM Cortex-A9
某些具有相同内核的 ARM SoC 是否比其他处理器更好？ （如果是这样，为什么？）
Is there much difference (in terms of performance) between two ARM SoCs that have the same core?
e.g., What differences (aside from those that can be attributed to the different clock rates) could be expected between:
- Apple A5, based on 1GHz dual-core ARM Cortex-A9
- Samsung Exynos GT-I9100, based on 1.2 GHz dual-core ARM Cortex-A9
Are some ARM SoCs with the same core better than others? (if so, why?)
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（2）
核心有转速数并且可以发挥作用。有时您可以使用 32 位 AXI 接口或 64 位 AXI 接口来获取它们，这对性能起着重要作用。另外，当然，对于 Cortex-A，您可能需要添加 L2 缓存，这将位于核心和需要花费一些时钟的东西之间。核心中有一些附加项目，例如 L2，我认为您不需要购买。
是的，您可以看到购买相同核心的供应商之间的差异。核心本身的性能可能与其家族成员或同名者非常相似。了解 ARM 停在核心或缓存或其他 ARM 附加组件的边缘。芯片供应商必须将 AXI 或 AMBA 总线转变为内部总线，并且在类似 Cortex-A 的情况下可能位于芯片外部（例如用于系统内存的 DRAM）。不同的芯片供应商的情况会有所不同，而且可能会有很大差异。例如，如果芯片支持 DDR DRAM，他们可能会从另一家供应商那里购买 IP 来实现 DDR 接口，并且每个芯片和芯片供应商都可以自由选择他们想要在其芯片中植入的任何 IP。
如果你留在核心，比如 L1 和 L2 内存系统，如果它有 L2，你可能会看到类似但不一定精确的性能（假设你以相同的时钟频率运行两个示例），一旦它超出手臂进入其余部分对于该芯片，我预计会有明显的性能差异。根据接口的不同，一个芯片可能会赢得一场性能竞赛，然后更换为另一个接口，另一个芯片可能会做得更好。
“更好的核心”这个词是一个充满争议的问题。您必须决定您认为“更好”的含义是什么。一种可能具有速度快的良好 I/O 组合，另一种可能具有良好的片上功能，但具有糟糕的外部接口。一个可能会提供一个真正的网络接口，另一个可能只会给你USB，你必须将其转换为网络、存储、视频等。除了你之外，其他任何人都无法回答你的这部分问题。如果您对产品有非常详细的需求规范，那么当然，社区可能会达成一致意见，但根据您的需求和竞赛功能的相似性，可能会导致平局。
我不知道你的项目/目标是什么，但如果你需要突破极限，你应该为每个参赛者购买或构建评估板，并将你的应用程序放在所有参赛者上，看看会发生什么。唯一重要的性能基准是您的应用程序在其他应用程序上运行的速度。
The cores have rev numbers and that plays a roll. And sometimes you can get them with say a 32 bit AXI interface or a 64 bit AXI interface and that plays a big role in performance. Also, certainly with a Cortex-A you probably want to add an L2 cache, and that will sit between the core and stuff costing you some clocks. there are add on items to the core like the L2 that I assume you are not required to buy.
Yes, you can see differences between vendors that have bought the same core. the core itself will likely perform very much like its family members or namesake. Understand that ARM stops at the edge of the core or the cache or other ARM add ons. the chip vendor has to turn that AXI or AMBA bus into something inside and in a case like a Cortex-A likely outside the chip (DRAM for system memory for example). And that will vary from chip vendor to chip vendor and can vary dramatically. If the chip supports DDR DRAM for example they probably bought a IP from yet another vendor to implement the DDR interface and each chip and chip vendor is free to chose whatever ip they want to plant in their chip.
If you stay in the core say L1 and L2 memory systems if it has an L2 you will probably see similar but not necessarily exact performance (assuming you ran your two examples at the same clock frequency), once it goes outside the arm into the rest of the chip all bets are off I would expect noticeable performance differences. Depending on the interface one chip would win one performance contest, then change to another interface and the other chip might do much better.
The term better core is a loaded question. You have to decide what you think is meant by "better". One might have a good mixture of I/O that is fast the other might have nice on chip stuff but have horrible external interfaces. One might provide a real network interface another might just give you usb and you have to turn that into network, storage, video, etc. That part of your question cannot be answered by anyone else but you. if you had a very detailed requirements spec for your product then sure, the community could converge on an opinion but depending on your requirements and similarities in features of the contest might result in a tie.
I have no idea what your project/goals are but if you are going to need to push then envelope you should buy or build eval boards for each of the contestants and put your application on all of them and see what happens. The only performance benchmark that matters is how fast YOUR application runs on something not some other application.
Snapdragon 在这里尤其不同。高通没有获得该架构的许可，而仅获得 ARM 的指令集许可，以保持与其他设备的二进制兼容。
他们声称它更像 A9 而不是 A8，但不幸的是 Snapdragon 属于主要 Coretex-A SoC 中最差的。
请注意，苹果的 A 系列和三星的 Hummingbird/Exynos 总体上是该联盟中表现最好的。
因此，如果您使用优化例程在 iPhone 上进行一些基准测试，请注意，它在具有不同 SoC 的其他平台上很可能表现不佳。差异可能很小，也可能很显着，具体取决于基准测试算法的性质。
此外，大多数与 MMU 相关的指令（例如 PLD/PLI）一直都是可选的，但对于 ARMv7，我不确定这一点。
干杯
Snapdragon is especially different here. Qualcomm didn't license the architecture, but only the instruction set from ARM just to remain binary compatible with others.
They claim it to be more like A9 than A8, but Snapdragon unfortunately belongs rather to the worse ones among major Coretex-A SoCs.
Note that Apple's A series and Samsung's Hummingbird/Exynos are overall the top performing ones in this league.
So, if you are doing some benchmarking on an iPhone with your optimized routines, beware that it will most probably perform not that great on other platforms with different SoCs. The difference might be marginal or significant, depending on the nature of the benchmarking algorithms.
Besides, most MMU related instructions like PLD/PLI have been optional all the time along, I don't know this for sure with the ARMv7 though.
cheers