2013年1月1日 星期二

Intel X86 及 ARM 都將走入一個超省電時代之CPU架構 ( Intel X86 and ARM will all goes into extra lower power multi-core CPU architecture )

英特爾點亮「非常多核心」未來  ( Intel lit massive multi-core future )

英特爾(Intel)稍早前釋出了‘Parallel JS’(平行JS),這是一個開放原始碼、數據平行版本的JavaScript,英特爾技術長 Justin Rattner 將之形容為非常多核心(many-core)運算發展過程中,向前跨出的一小步。
Rattner稍早前在英特爾開發者論壇(IDF)發表主題演講,並針對平行編程和降低PC、伺服器等功耗,展示了最新的程式語言和其他研究成果。

包括英特爾、微軟(Microsoft)、Nvidia和其他業者,都對大學研究單位挹注了大量投資,以定義能滿足未來非常多核心處理器需求的編程工具。截至目前,平行化編程一直僅被應用在高度專業化的科技應用領域。

“我們已取得良好進展,但未來將不會只有單一[編程]模型,而是會有許多個模型,”Rattner說。Parallel JS是其中一個代表模型。這種語言針對數據密集型的運算、基於瀏覽器的應用如照片和視訊編輯,以及3D遊戲等在英特爾晶片上執行的應用程式提高性能。其主要目標是吸引使用腳本語言的主流Web程式設計師們。

Rattner還展示了該語言可在高階動畫應用中處理英特爾CPU上8個x86核心的能力。

“最近一段時間以來,大多數的軟體都是用Java或Python等腳本語言編寫,但至今這些程式設計師還未真正使用多核心工具,”Rattner說。Parallel JS是“非常重要的一步,它讓我們超越了傳統思考侷限,一旦你突破了少數幾個核心,那麼,建構多核心晶片就只是應用程式的技術罷了,”他說。

該語言的未來版本也將充分利用目前嵌入在英特爾最新處理器中的繪圖核心。為此,Rattner展示了用於x86和繪圖核心的臉部辨識應用程式。

“我們基本上是告訴開發者,是時候去思考異質運算的創造性了,”Rattner說。

非常多核心和行動願景

在實驗室中,英特爾也正在研究如何改善當前在繪圖處理器上用於執行通用程式的數據平行工具。包括今天的 OpenCL和Nvidia的Cuda工具等,都使用與硬體緊密聯繫、相對較低階的數據原件(data primitives),Rattner說。

英特爾正在使用更高層的編程抽象,如用於密集和稀疏矩陣算法中的嵌套向量(nested vectors)。Rattner稱該公司可能在2012年發佈這些工具。

新的軟體代表了一些學術研究單位致力於將所謂的‘函數編程’(functional programming)概念帶進今天的C++語言中。

“函數編程看起來是平行編程朝更高階抽象與更加自動化和並行方向發展的基礎之一,”Rattner說。“編譯器可以萃取並行,而且程式設計師毋須再像使用OpenCL或Cuda做陳述,”他表示。

2012年以後,數據平行技術將產生根本上的變化,Rattner說。今天,每處理一次任務都是在嚴格調度情況下進行,但這會讓部份電腦資源閒置,造成能源的浪費。

而未來的方法基本上會是非同步處理,但目前僅在概念階段。“今天,為了讓編程更加便利,我們放棄了效率,但展望未來,我們沒有理由再浪費這麼多的電力了,”他說。

節省PC耗電

Rattner展示了兩項專門針對減少運算功耗而進行的研究專案進展。

一個近閾值電壓處理器採用嶄新的低電壓電路,其運作接近閾值水準。其概念是當需要時CPU的運作速度必須足夠快,但當下降到低於10毫瓦功率時,便會處於輕載模式。為了展示這個概念,英特爾開發了一款能以郵票尺寸的太陽能電池運作的Pentium級晶片。

這款展示晶片命名為Claremont,僅執行在比閾值電壓高100毫伏的電壓水準,與現有處理器相比功耗可減少5~10倍。“這是一個龐大的數字──人們長久以來一直為爭取減少20%功耗而努力,所以這幾乎是前所未聞的成就,”Rattner說。

Claremont僅使用一個L1快取,因為相關記憶體仍需運作在閾值以上的數百毫伏水準,Rattner說。此外,英特爾和美光(Micron)的研究人員還發現了一種新型堆疊技術,並共同開發了新的記憶體原型。這個混合式記憶體立方體(Hybrid Memory Cube)結合了頂端的DRAM晶粒堆疊和底部邏輯層,採用了全新介面和協議,以便將記憶體資訊轉換到獨立處理器。

英特爾研究員Bryan Casper表示,“當測量位元傳輸量與能耗時,該元件是有史以來能效最高的DRAM。”與目前最先進的DDR4記憶體模組相比,這個原型的頻寬高出10倍,能效則高出7倍,他表示。

此外,Rattner還展示了可作為基地台的標準x86伺服器,它在Sandy Bridge CPU上採用新的x86訊號處理演算法。

“我想,我們身在一個可以開發出某些真正有價值的東西的時代,”他說。“對一些傳統上一直採用以DSP為主的系統而言,我認為Ivy Bridge或許具備能與之競爭的潛力,”他表示。

在本屆IDF上,英特爾的工程師還針對其處理器講解了新的訊號處理和封包處理開發套件。

ARM unveils chip designs for future smartphones, servers

英國芯片設計商ARM週二推出的處理器電源新一代的智能手機,而且還提供低能源解決方案的服務器,增加侵入英特爾主導的市場機會。British chip designer ARM on Tuesday launched processors to power a new generation of smartphones and also offer low-energy solutions for servers, increasing its incursion into a market dominated by Intel.

這劍橋的公司,其技術是在蘋果公司的iPhone 5和三星電子的Galaxy S3顯示出實力,其最新的藍圖,將提供3倍的處理能力,使用相同數量的電量。The Cambridge-based company, whose technology is in Apple's iPhone 5 and Samsung Electronics' Galaxy S3, said its latest blueprints would deliver three times today's processing power using the same amount of energy.

The new chips use 64-bit architecture, an upgrade from current 32-bit designs that will give them increased processing clout suitable for servers while retaining their energy efficiency, the company said.

該公司表示,這種新的芯片使用64位體系結構,從當前的32位設計升級,這將使他們增加適合的服務器的處理的影響力,同時保留其能源效率。Chips made using technology licensed from ARM are used across the smartphone and tablet industry. The company is now promoting an emerging trend toward building data centers with many small, low-energy chips instead of a few high-performance ones.

微服務器行業正處於起步階段,惠普,戴爾等廠商表現出極大的興趣。The microserver industry is in its infancy, with Hewlett-Packard, Dell and other vendors showing keen interest.

PC芯片製造商 AMD 週一表示,將使用ARM的技術,使微服務器芯片在2014年開始。PC chipmaker Advanced Micro Devices said on Monday it would use ARM's technology to make microserver chips available starting in 2014.

ARM Chief Executive Warren East told reporters at an event in San Francisco that ARM-based servers may account for a fifth of data centers by 2020.

他說興趣微型服務器蔓延到最先進的互聯網公司如 Facebook 和谷歌傳統保守的客戶,如銀行的首席信主管。He said interest in microservers is spreading beyond cutting-edge Internet companies like Facebook and Google to traditionally conservative customers, like banks' chief information officers.

"Getting to the CIOs of banks is quite a step but we're already seeing these types of people totally plugged into the notion of ARM-based servers, and looking at not so much whether, but how and when," East said.

這新介紹64位技術,可以讓處理器的互動,更有效地使用內存,已經由英特爾提供的一個功能,是ARM芯片為數據中心更具吸引力的關鍵一步。Introducing 64-bit technology that lets processors interact more efficiently with memory chips, a feature already offered by Intel, is a key step toward making ARM chips more attractive for data centers.

East 說,位於得克薩斯州奧斯汀的啟動 Calxeda 已經證明,降低超過四分之三的服務器功耗的ARM技術。East said Austin, Texas-startup Calxeda has demonstrated that ARM technology can reduce power consumption by over three quarters in servers.

Unlike in the PC industry, which Intel dominated for decades, manufacturers in the smartphone industry can buy their processors from any number of chip designers.

East said data-center equipment manufacturers will also benefit from more options as ARM-based chip designers launch competing chips into the server market, long an Intel stronghold.

Licensees of ARM's new 64-bit Cortex A-50 series include AMD, Broadcom, Calxeda, HiSilicon, Samsung and STMicroelectronics, the company said, and the first chips are expected to ship in 2014.

AMD: ARM's power advantages could wane in the coming years ( AMD:ARM的省電優勢可能在未來幾年會減弱 )

ARM's power advantages over typical x86 chips from Intel and AMD will continue for around half a decade before fading away, according to AMD.Yesterday, AMD announced it would start making ARM-based servers in 2014, and today Suresh Gopalakrishnan, general manager of AMD's server business unit, told ZDNet that ARM's power advantages could be short-lived.

"If you increase the capabilities of the [ARM] processor, it will consume more power. At the same time, if you start working on the [x86] architecture and process together you can bring the power down as well," Gopalakrishnan told me. "Right now ARM has the [power] advantage - five years down the road we'll see."

在未來,英特爾將帶來新節能,節儉芯片建立在其先進的14nm和10nm的製造方法,這將有助於它降低其功耗,而ARM的芯片將消耗更多的功率為更多的功能被加入到他們讓他們解決更高級的服務器工作負載。In the future, Intel will bring in new energy-thrifty chips built on its advanced 14nm and 10nm fabrication methods, which will help it lower their power consumption, while ARM's chips will consume more power as more features are added to them to let them tackle more advanced server workloads.

So, in a few years, the power advantages which are propelling ARM's chips into cloud datacentres like Facebook's could fade as technological forces conspire to close the gap in electricity usage.

Why is AMD licensing ARM for servers now then? Because it can bring the 64-bit chips out years before the power window closes, Gopalakrishnan said.

Along with this, by pairing ARM chips with a dense networking fabric, like AMD SeaMicro's Freedom Fabric, there's a chance to create some novel server designs: "You will see very dense clusters tuned towards a certain class of workloads," Gopalkrishnan said. Some of these workloads will be ARM and some will be AMD.

As AMD says, the future is 'ambidextrous' and from 2014 it's going to have an x86 hand and an ARM hand. However, for how long this will remain is a mystery if the power advantages disappear.

分析

  • Intel 將在 2013 Q4 ~ 2014 面臨 ARM 64bit 晶片之競爭,畢竟,ARM 64bit 晶片能提供 cost effective 及低功耗之能力超越 Intel 晶片;
  • ARM 64bit 晶片之決勝關鍵仍決定於像 TSMC Samsung 它們的製程與 Intel 差距,目前看來最大優勢期間是自 2013 Q3 ~ 2014 Q4,2015 開始 Intel 還會再度領先;
  • Calxeda已經推出32位元處理器EnergyCore,而隨著ARM公布了64位元ARMv8架構,預期該公司也會繼續針對伺服器市場發表新產品。
  • 台廠要有加速運用 ARM CPU 計劃,以避免 Intel 失去地位時引來 Samsung 競爭之大衰退;
Enhanced by Zemanta

沒有留言: