標題: 雙核 A15 大戰雙核 Atom Benchmark（Ubuntu） [打印本頁]

---

作者: liangsuilong    時間: 2012-11-22 00:16     標題: 雙核 A15 大戰雙核 Atom Benchmark（Ubuntu）

Benchmark 原文係哩度：
http://www.phoronix.com/vr.php?view=18159

Phoronix 將部 Samsung Chromebook 整左個 Ubuntu 13.04 for ARM 落去，同 Atom D525 鬥左下，加埋粒 A9 四核對比左下。哩個 Benchmark A15 雙核係 Compiler 同埋一D Software 會留左力，預設 FPU 係行 VFPv3，A15 可以用新一代嘅 VFPv4。睇落哩幾個測試 A15 同 Atom 差唔多咁快，但係如果 Software 針對得多核心優化得好D，A15 雙核依然會輸畀 A9 四核。

我嘅評論：Atom 想係 2013 發圍難度都好大，新一代用 22nm LP 嘅 Atom 要到 2014 年年頭先出到。2013 年 Tablet 同高階 Smartphone 主打都係四核心 A15（Krait），中階會係以新製程高頻 A9 或雙核 A15（Krait） 主打，加上 Atom SoC 果粒 GPU 歷來都唔算特別勁。Intel 嘅 Mobile 之路仲係好長。但係我真係唔明 Intel 22nm LP 點解要拖咁耐？

---

作者: toylet    時間: 2012-11-22 00:50

提示: 作者被禁止或刪除 內容自動屏蔽

---

作者: singshan    時間: 2012-11-22 01:02

X86 喺發熱上係比 arm 難控制好多

---

作者: singshan    時間: 2012-11-22 01:05




    仲有用電量

---

作者: 小安哥    時間: 2012-11-22 01:05




    next gen atom soc uses hd4000 and have a very low power consumption, it makes itself power and performance competitive to those ARMs cpu, and it supports the x-86, which is great.
ARMs must have some new product other than big little to win intel this time, as next gen atom also have a 40% performance gain.

---

作者: cksin    時間: 2012-11-22 01:13




    咁你又知吾知點解呀?

---

作者: marcochl    時間: 2012-11-22 01:14

回復 4# singshan

都要睇下係邊粒x86, 邊粒arm...
S4 Quad  既power consumption/heat 已經唔低...
另外intel 其實係有D做過手腳既atom....果粒先係真正打ARM 用既...
個test 用果粒係std atom....max TDP13W...

呢D browser test 只可以證明arm 係個別情況(Chromebook) 下run 某D 野已經追到x86
而仲未可以代表arm 可以ko x86...

---

作者: cksin    時間: 2012-11-22 01:21




    到時如果Atom table 用到PhotoShop CS5, 甚至 LOL遊戲, 吾知你又會點睇.
係因為ARM 架構比較簡單 開發 製做都比較快, trade off 係功能上比X86少好多.Atom 點解睇落去個spec. 吾係好勁, 係因為要花比較多既時間係開發方面

---

作者: cksin    時間: 2012-11-22 01:28




    無錯, 嘩眾取寵既一個chromebook test, 實際完全冇意思
好似叫F1賽車同普通私家車放係旺角繁忙時間鬥快咁

---

作者: cwkan    時間: 2012-11-22 02:06

好似幾吸引喎

---

作者: peaceso    時間: 2012-11-22 11:24



Q 記一向大食. 因為是它自己設計
要看S, T, A, N 記的U

---

作者: peaceso    時間: 2012-11-22 11:26




  Atom 一向差....P4 Architecture

---

作者: marcochl    時間: 2012-11-22 11:46

回復 11# peaceso

Eyxnos 5系已經開始又熱又用多電
N 記就可以節省d, communication module 係死症
至於其他兩間唔清楚
但係TSMC 28工藝…………

---

作者: waikhplkcc    時間: 2012-11-22 11:52

TI 既A15 依家都未出, 真係慢到HIHI.
Samsung 就一向用電多, 熱量大, 都係HIHI.

---

作者: willy930    時間: 2012-11-22 12:39



Atom is based on Pentium MMX.

That's why it is still not an OOO processor.

---

作者: peaceso    時間: 2012-11-22 14:03

Pentium 1, 2, 3 are similar architecture.
Please advice why first gen atom is poor performance when compare with 1GHz Pentium 3.....
Atom is base on P4 which is mentioned by Intel.
Furthermore, P4 is also come from Pentium MMX....
While there is two type of Atom... from the report, it is P4 base.

---

作者: willy930    時間: 2012-11-22 15:43

本帖最後由 willy930 於 2012-11-22 15:45 編輯


P4 is OOO already. Atom is not. This alone mean that they are two different designs. It takes a lot of re-design to change an in-order execution pipeline to an OOO execution pipeline.

P4 is based on NetBurst architecture. It does not share a lot with P6 micro-architecture which powered Pentium Pro/Pentium II/Pentium III. The original Pentium and Pentium MMX are based on P5 architecture and do not support OOO execution. Atom are based on Bonnell micro-architecture, which shares a lot of similarities with the Pentium/Pentium MMX.

P6 micro-architecture is not really based on P5 micro-architecture. And NetBrust is not really based on P6 micro-architecture.



http://en.wikipedia.org/wiki/Pentium_4



http://en.wikipedia.org/wiki/P6_%28microarchitecture%29



http://en.wikipedia.org/wiki/NetBurst_%28microarchitecture%29



http://en.wikipedia.org/wiki/Bonnell_%28microarchitecture%29




http://www.anandtech.com/show/2449/2

The ability of doing OOO execution can increase performance by a lot already. This is something the Atom family does not have. OOO execution means that code are executed in term of data flow. If the data is not there yet, OOO execution design can switch to handle another thread first. In-order execution means that pipeline will follow the order of the instruction code does not care whether the data is there yet. So when the data is not there yet, you are stuck.

Also, Atom has only 2 ALU and maximum 2 instruction decode rate does not help its performance also, but Intel can not put more units as they want Atom to be a power efficient processor. Furthermore, Atom has Hyperthreading does not mean it is based on Pentium IV. It is just borrowing useful pieces.

---

作者: snoopy11hk    時間: 2012-11-22 17:09




    TI 唔係放棄左 ARM 呢個市場?
我見佢連 A50 series 都無玩囉喎

---

作者: willy930    時間: 2012-11-22 19:00



Ti gave up ARM for mobile device market.

They will continue to develope ARM SOC for other purposes.

---

作者: snoopy11hk    時間: 2012-11-22 19:14




    無 A50 series 連之後的 ARM server 都無得玩

---

作者: Puff    時間: 2012-11-22 20:15


Moving away from mobile devices market is not the same as not licensing Cortex A50 cores.

---

作者: Puff    時間: 2012-11-22 20:18

本帖最後由 Puff 於 2012-11-22 20:20 編輯

Because this is how the development pipeline works. They have already accelerated the roadmap, and the effect will start showing up in 2013 and 2014, with moving to two smaller process nodes in only two years, catching up with the process roadmap of the PC Haswell line.

---

作者: Puff    時間: 2012-11-22 20:24


... while the journalists generally think that Krait and Swift are the cores designed with the low-power mind, as compared to the said-to-be-big A15.

---

作者: KwokAngus    時間: 2012-11-22 21:06




    Well, I personally agree with it.
But to verify the claim, we need a 28nm-Dual A9
otherwise it is not fair to 40nm-Dual A9

---

作者: cksin    時間: 2012-11-22 21:47




     Q記將無線電晶片和電源處理晶片都放入係同一粒SoC ,依個先係你所謂高通自己設計, 內核則依賴Acorn Computers提供ARM core design

---

作者: KwokAngus    時間: 2012-11-22 21:52




    噢
atom 1C2T 用電量同 dual A9 差不多
但前者快過後者

計番條數單核in of order+bonwell atom core咁落後嘅microarchitecture都咁快, 不得了...

---

作者: Puff    時間: 2012-11-22 22:02

本帖最後由 Puff 於 2012-11-22 22:05 編輯

Qualcomm licenses the ARM ISA and does their own core design. At least it is not promoted as an ARM offering but a custom core in the public.

---

作者: KwokAngus    時間: 2012-11-22 23:00




    其實唔係x86比ARM 難控制 heat dissipation, power consumption
而係設計, x86以往設計上, 架構取向一直以效能為主
x86一樣可以有低電低熱嘅core

---

作者: KwokAngus    時間: 2012-11-22 23:05

本帖最後由 KwokAngus 於 2012-11-22 23:11 編輯



    um... yes.. just my logic head asks me being aware conventional dual A9 cores of 40nm
Krait and Swift are really good

And people are really not so realistic to imagine dual A15 could save many power, or being very power efficient than dual A9
IN fact, A15 itself is performance-oriented design. With aid of other technologies, the final product might be good. But not A15 itself, I believe.
Cortex A7 is the Mr. Right for power-oriented design.

I don't know whether it is called power-oriented or efficiency-oriented.

---

作者: willy930    時間: 2012-11-22 23:31



If I remember correctly, I believe AnandTech said that A15 has more deapth as it has a longer pipeline to go for higer clock and Krait/swift are wider.

I think ARM is really pushing the big.LITTLE thingy. When running background/OS processes, use the LITTLE cores. When the user start an app, switch over to the big cores.

---

作者: 實頭    時間: 2012-11-23 00:24




krait 係鬼A15

---

作者: liangsuilong    時間: 2012-11-23 12:01



類 A15 嘛。。

---

作者: liangsuilong    時間: 2012-11-23 12:05



22nm Atom 要 2014 年先出到，咁 TSMC 果邊 20nm 都開始量產，雖然佢地一開始會超多問題。Atom 個市場前景都唔係咁樂觀。Boost Roadmap 我估起碼要到 2015 先開始有效果。2015 都開始有第二代 ARMv8 啦。

---

作者: liangsuilong    時間: 2012-11-23 12:11



I am not worried about performance of next gen Atom. I just worried about when Intel rolls out it and how it competes with other ARM chips. Time is as important as technology.

---

作者: liangsuilong    時間: 2012-11-23 12:23



1. ARM 嘅效能足夠，且用家足夠多嘅話，軟件商唔會坐視哩個市場唔理。Adobe 已經出左 Photoshop Touch 畀 iOS 同 Android，EA 亦有唔少 Game Port 左落 ARM，最新一代嘅 NFS，講左好耐都係未做好嘅 Battlefield 3.。Microsoft Office 都有 ARM 版。

2. 好似哩D咁嘅 Form Factor，着重玩同輕量使用。試問是否適合放一個 Full Function 嘅 Photoshop 或其他專業級生產力工具落去呢？我唔覺得 Atom Tablet 行 Photoshop CS5 會有咩好體驗，就算唔計效能只講 UI。

3. 軟件功能可以做到幾多嘢，好睇個硬件效能去到邊度，同埋個 OS 做到咩程度。ARM 同 x86 用同一個 OS 嘅話，效能又差唔多，App 又有齊嘅話，我唔覺得 ARM replace 唔到 x86 咯。

4. 原文只不過講左話 ARM 都可以做出一粒效能追得上入門級 x86 嘅 CPU 姐，唔好講到 CPU 效能以外嘅嘢咯。

---

作者: willy930    時間: 2012-11-23 13:11



I don't think NEON is anywhere close to the FPU on Atom/Bobcat.

Later on, when mobile apps use more floating points, ARM may have a hard time.

Of cause, there is still something call OpenCL which may be able to help the ARM camp.

---

作者: griffinlam    時間: 2012-11-23 14:08

提示: 作者被禁止或刪除 內容自動屏蔽

---

作者: Puff    時間: 2012-11-23 17:49


下代 22nm Atom，我睇到既資訊都依然係話 13H2 推出市場。

而代工廠既 20nm 係 13H2 只係 ramping up. 只係 ramping up. 按 40nm/28nm 既往續黎睇，無返一年半載係無 mobile SoC vendor 會 launch new products，產能既供給水平同可能出現既良率問題。而 Intel 既 22nm 已經係 well-established，更加唔好講 14H2 佢地就預計會推出 14nm Atom.出唔出到另一個問題。

---

作者: Ksec    時間: 2012-11-23 18:22



TSMC 20nm production will be coming in much quicker then you think then. Not to mention they have an extra 1B R&D left over from budget for improving 28nm.

You are very likely to see TSMC 20nm SoC product on market in Q3 / Q4 2013.

---

作者: Puff    時間: 2012-11-23 20:29

本帖最後由 Puff 於 2012-11-23 20:30 編輯

Perhaps. Will see FPGA at first anyway.  

---

作者: liangsuilong    時間: 2012-11-25 22:52

本帖最後由 liangsuilong 於 2012-11-25 22:56 編輯


咁就唔知道喇，之前有條新聞話 Tablet 版嘅 22nm Bay Trail 要到 2014H1 先出貨。但係我就覺得好奇怪咯。22nm 已經出左好耐，LP 工藝冇理由仲要搞咁耐啵，而且咁遲出四核 A15 都收尾等 A57 四核出，到時出境真係好難打。

TSMC 20nm 我覺得佢地真係可以係 13Q4 出到少量貨畀 HP 嘅 A/N Desktop GPU 同埋 Mobile SoC 用，前提係佢地真係可以按照原定計劃順利進行。問題就係佢地通常都跟唔上計劃而且技術上通常有唔多順利。哈哈哈！

---

作者: cksin    時間: 2012-11-25 23:50




    1. 本身 ARM 設計上就吾係用黎做desktop做到既野!  你見到而家PS cs5都已經可以推到部desktop(約300w, 125wTDP,4GBram)好盡, 點解係ps touch 又得? 咁係因應ARM 本身有限的效能PS touch 本身亦都減少左好多功能 同D手機影相apps 相差冇幾.

2. 我為左回你話"2013 年 Tablet 同高階 Smartphone 主打都係四核心 A15（Krait），中階會係以新製程高頻 A9 或雙核 A15（Krait） 主打，加上 Atom SoC 果粒 GPU 歷來都唔算特別勁。Intel 嘅 Mobile 之路仲係好長"  其實如果係五年前, 好多人都會覺得而家既智能手機冇咩用, "邊有人用手機上網做野嫁,返公司做啦, 咁細個mon 上網又慢""3吋mon 56kb 上網咁慢, 用電好似倒水咁,點用黎睇片呀" 你今時今日有冇林過其實幾年前冇咩可能既野, 今日通通都做到. 其實個問題係睇下係件電子產品比到幾多既功能出黎既姐, 有用既功能,人係會改變自己既使用習慣

3. oh, sorry 原來你連OS係咩意思都吾知...

4. 我正正係對文中,咁樣既比較方法好不滿, 有好大既誤導成份, 令到好多對電腦一知半解既人以為ARM 都可以做出一粒效能追得上入門級 x86 嘅 CPU

---

作者: cksin    時間: 2012-11-25 23:54




    哦原來係咁, 但係我覺得佢地吾係將效能放係睇一位, 而係將SOC 放係第一位, 務求減輕下級生產廠成本

---

作者: liangsuilong    時間: 2012-11-26 00:07



http://www.phoronix.com/scan.php ... exa15_gcc&num=1

Phoronix releases more results on Samsung Exynos 5 Dual. New GCC could improve the performance.

Another guy releases the result comparing with AMD E-450:
http://openbenchmarking.org/result/1211243-SU-1211246RA29

Exynos has 128bit DDR3 memory controller, more bandwidth than Atom and Bobcat. I think Cortex-A15 cannot beat Atom and Bobcat in FPU part now. But it is following up. Software still stay tuned for Cortex-A15. As we know, all the Linux benchmark is disabling VFPv4 support at this time. I could not know how much improvement from VFPv3 to VFPv4. In the past, Hardfp is also disabled by default on ARM benchmark, then we can see FPU performance is very terrible on ARM.

The story is more and more interesting. Let's see it in the future.

I am quite interested in ARM Green HPC. A SoC could  contain 16 ARM Cores in the chip, 4 Clusters, 4 Cores in a cluster.

---

作者: willy930    時間: 2012-11-26 10:26



The test kinda proved that Exynos 5 does have a better cache/memory system.

And ARM is finally getting very close to a real x86 CPU cores. So +1 for them.

But, don't forget that bobcat only has a 64 bit FPU. Jaguar will be the real test for Exynos 5.

And next Atom will be partially OOO too. So the battle for low power CPUs has just started.

---

作者: Puff    時間: 2012-11-27 00:28


Yep. Power efficiency and their own SoC roadmap at first. Performance at second.

---

作者: 86player    時間: 2012-11-27 22:37

ARM百花齊放 INTEL雄風難再現於移動領域
INTEL還是專心搞好桌面電腦吧

---

作者: liangsuilong    時間: 2012-11-30 00:10



http://www.phoronix.com/scan.php ... nos5_dual&num=1

New benchmark continues. Quad Core Cortex-A15 will be mainstream in the next year. I think it could follow up the the first generation low-end Core 2 Duo in multi-threading benchmark, at least Conroe-based Pentium Dual-Core. The story just starts at this time.

Intel maybe feels worried whether OoO Dual-Core Atom for Tablet can catch up ARM or not. I just hope new Atom will releases on time, not delay to 2014H1. I still believe OOO Atom has a great improvement comparing to in-order Atom. 40% or more? I am not sure.

AMD Jaguar core is the same. Jaguar is based on Bulldozer design. I think single core of Jaguar is not fast than Cortex-A15. Maybe one Jaguar module is equal to 1.5X Cortex-A15 core. We should not only focus on performance. Time is more important than performance.

---

作者: willy930    時間: 2012-11-30 10:41




The next Atom will not be full OOO. It will be partially OOO. However, it will support all the new instruction set. Too early to judge its improvement over the previous generation of Atom.

Jaguar is not based on Bulldozer. Jaguar is more like a super tweaked bobcat @ 28nm.

- New L2 cache interface
- 128 bit FP Unit (Double pumped to support AVX-256)
- Enhanced Integer pipeline
- Higher clock
- 28nm

~= 15% improvement in general.

Next Atom will be release in 2H 2013. Jaguar will most likely be demoed during CES 2013 (January).

---

作者: liangsuilong    時間: 2012-12-2 02:41



Could you talk more details about next Atom? I do not know understand how it works under partially OOO.

And Jaguar shares all L2 cache with four cores and has two 128bit FPU for per core? I believed Jaguar just have flexible FPU design like Bulldozer before. Now I should change my mind. I just hope AMD should not only focus on Windows Tablet. I hope Temash will ship as soon as possible and support Android.

---

作者: willy930    時間: 2012-12-3 10:59



Sorry. I do not have any more information other than knowing that the next ATOM will be partially OOO. Don't know how Intel is going to archive this.

Each Jaguar core will have one 128 bit FPU. Shared L2 for all versions. Temash should have Linux support. Hondo is too late to the market. No point to implement the linux stack for it.

---

作者: anonymoushk    時間: 2012-12-3 11:17


x2
我地電腦用果D ps / lightroom … 一張raw都 21MB，仲有一大堆layer 同 effect 要放上去，電話版邊有得比 xddddddd
   
相越大越食ram…

---

作者: cksin    時間: 2012-12-4 00:55




    如果有玩過Cortex-A8/9 /M4 就會知ARM行X86係一件冇可能既事, 得到既效能會令你恍惚置身於十年前
相比之下, D RAM 真係冇得比.

---

作者: liangsuilong    時間: 2012-12-4 01:23



我林我同你之間嘅理解有小小偏差咯。

1. 真係要執相嘅一定要用部 Powerful x86 PC，唔勁都唔用，靠佢搵食嘅人就更加唔使講，哩個我十分同意。但係一般人更多嘅係用平板或手機加個 Effect/Filter 加個相框畀幅相 Send 上網，咁係米要需要一部好勁嘅 PC 執相。我認爲係唔需要咯。做哩D嘢梗係用 PS Touch 好過全功能版 Photoshop 方便兼好用好多啦。

2. 以後發展成點唔知，但係 Intel 就一定唔會畀 Atom 行得勁，唔係會影響到 i3/i5/i7 嘅銷量，哩個係 Intel 產品定位嘅問題，而唔係佢以後會發展到好快做唔做到嘅問題。所以你宜家開全功能版 Photoshop 執相一定唔會買部 Atom 機，最少都會買部 i3 或者 AMD 同級嘅產品。你宜家出街影左相加個 Effect 發上網都會係用翻 ARM 嘅手機或者平板，因爲 Atom 唔夠慳電。就咁簡單！當然 Intel 屈機到將 i7 改成用 Atom 品牌來出咁我都冇嘢講。

3. OK，之前我哩段話講到唔好，我收翻！我係話 ARM 機越出越多，軟件公司一定會睇到哩個潮流，幫佢寫 OS 寫軟件兼且做優化，不過哩個唔係一蹴而就，係一個循序漸進嘅過程。舊嘅 PC Game 都有幾隻 Port 左落 iOS 同 Android，又有唔少公司出做佢地軟件功能縮水嘅 iOS 同 Android 版，微軟都搞左個 Windows RT 雖然用落就廢廢地。

4. 硬件效能上講，Cortex-A15 已經追得上 Atom 同 AMD Bobcat 哩類入門級 x86，只不過傳統 Windows Software 本身冇 ARM 版所以先冇 App 用。未來會慢慢改變，特別 Windows RT 出左之後。用 Linux 做 Review 係因爲 Linux 好多軟件都係開放源代碼，可以係 ARM 上面行。但 Mobile OS 絕大部分都有爲 ARM 做優化，比如 iOS 同 Android。Mobile Device 嘅趨勢又一定程度上係日常使用逐漸替代 x86 PC（注意：不是完全替代）。所以 ARM 先會有咁多人關注。

---

作者: willy930    時間: 2012-12-4 10:20



Re your point #4, we have not seen the true potential of x86 when running on mobile OS yet. A lot of mobile OS are not optimized for x86 yet. Is true that x86 CPUs are still going to be more power hungry. But, its performance/watt should go up when people start to tweak for x86.

Make no mistake, A15 is good, but at the same time, x86 can be better.

---

作者: cksin    時間: 2012-12-5 02:23

本帖最後由 cksin 於 2012-12-5 02:29 編輯


嗯 我林總括而言 你對成個技術,市場既認識同定位都未有一個solid 既觀念, 單單用效能係冇法有效評估一個產品
例如Cortex M3同 M4同A8 既分別... 同一個家族 但係M4多左個floating point unit
M4 頻率比M3多左一倍, 但係效能係吾係多左一倍? 其實遠遠多過一倍,因為佢多左個floating point unit 咁但係會吾會有一日M4 會取代到A8, 答案係絕對吾會, 因為係MMU結構上, 係兩回事, A8 行到linux os 而M4係冇依個本錢去行
或者你吾係依行,吾知咁多都好正常
有好多時吾係追吾追上, 而係本身個目標吾同, 但萬一個目標市場重疊了, 兩者一定會試去拉開個差別, 你要知道除非你有壓倒一切既表現, 否則冇人會去直接同另一個市場內的產品爭天下