What changes with "Snapdragon 845"?...
15
01
What changes with "Snapdragon 845"?Explain the contents of "inherent evolution": Snapdragon Tech Summit 2017 (page 1/3)
As mentioned in the previous report, the name of the next flagship processor of Qualcomm's next flagship processor is "Snapdragon 845", but this time I will look a little more in detail about the contents.
Snapdragon 845を構成する機能群In a nutshell, Snapdragon 845 is a "inherent evolution".
One year after the previous model's 835 announcement, functional functions have been gradually strengthened.The CPU core is from "Kryo 280" in the 835 era to "Kryo 385", 4x4 BIG.The point of the Little configuration has not changed.The maximum operation clock frequency of the BIG side core is 2.It is 8GHz, up 25-30 % from the previous model, the maximum operation clock frequency of the Little side core is 1.8GHz will increase 15 %.
Which core is used is determined by the software side, so the performance and power saving at normal use depend on how the OS is tuned.However, in terms of peak performance, it can be considered that there are about 30 % improvement compared to 835.
Snapdragon 845はSamsungの第2世代10nm LPP FinFETを使って製造され、その中心となるKryo 385はARM Cortex-Aベースのカスタムコアとなっている。基本的な構成は835時代から大きく変化していないが、全体にパフォーマンスが向上しているAnd the biggest difference from the previous model in 845 is the L3 cache.Until the L2 cache is assigned to a core, the 2MB L3 cache is shared in all cores.In the 835, the BIG core had a 2MB L2 cache and the Little core had a 1MB L2 cache, but in 845, the size was unknown, but the L3 cache was prepared while leaving the L2 cache.
The role of the cache is to reduce the access to the main memory and improve the response, but simply assigning the appropriate capacity cache for each core seems more efficient.However, in fact, multiple threads will be assigned to various cores including BIG and Little according to the operating status, so it is better to share the necessary information between the core and cooperate.Is easy to improve is the concept of cash sharing using Dynamiq adopted in Kryo 385.
In recent years, there is a tendency to adopt a mechanism of "heterogenia score" that enhances individual processing and power -saving, with recent processors in -house, but that trend is even stronger every year.。Snapdragon 845 is equipped with CPU core Kryo 385, GPU core Adreno 630, DSP HEXAGON 685, and SPECTRA 280 ISP of sensor processing unit.ADRENO is further subdivided, including a dedicated unit specializing in video encoding and decoding processing, and constitutes Snapdragon in a combination of all of these.
In addition, "system cache" is assigned about 3MB for the entire Snapdragon 845.Unlike the KRYO 835 L3 cache mentioned above, this is an optimization of performance by reducing memory access to the entire processor.Although it is a structure that reflects the nature of the heterogenia score, there is a device that aims at the performance boost without reducing power saving.
近年のプロセッサの特徴として、機能別に専用のコアや機能ブロックを用意するというヘテロジニアス化がより進んでおり、電力効率面やパフォーマンス面で有利になっているWindows on Snapdragon 835のデバイスでタスクマネージャーを開いたところ。機能別にコアの稼働状況が確認できるシステムメモリと呼ばれるメインメモリとキャッシュを仲介する領域がプロセッサに確保されている。メインメモリアクセスを40〜75％ほど低減し、ボトルネックを解消するというAnother trend in the processor includes a built -in processor core, such as Apple's A11 Bionic and Huawei's Kirin 970, optimized for so -called AI deep learning processing.Although the response of each chip vendor company for this trend is different, the calculation is high for Snapdragon by combining Kryo + Adreno + Hexagon and vector operative units.
Although this structure in 845 does not change from 835, the performance of individual core is improved overall, and the functions of Qualcomm are expected to improve the functions of three times the conventional.It is also characterized by supporting various DNN libraries such as TensorFlow, TensorFlow Lite, Caffe, Caffe2, Onnx (DEEP NEURAL Network) library, and this area is a general -purpose (Hanyo) processor assuming that various end users use it.It can be said that it is unique.
話題のAI処理については、AppleのA11 BionicやHuawei Kirin 970のNPUのような専用のコプロセッサユニットを搭載するのではなく、CPU＋GPUのほか、Hexagon 685 DSPのようなベクトル演算ユニットを組み合わせてディープラーニング系の処理を行う。これら演算ユニットは従来と比べてそれぞれ強化されており、Snapdragon 835との比較で3倍の高速化が可能としているソフトウェアベースの処理が中心になる代わりにTensorFlowをはじめとする各種DNNライブラリをサポートし、この点での汎用性を高めているディープラーニング向けライブラリでも利用されているベクトル演算DSPのHexagonもプロセッサの世代ごとに進化しているHexagonをサポートする周辺ハードウェアやサービスのエコシステムUltraHD Premiumコンテンツをキャプチャーできる1|2|3次のページへ
