星期二, 7月 19, 2005

IBM Power6的可能規格

IBM這次的口風真的是有夠緊,到現在連個code name都不知道。不過,情報還是會自己找出路的。

消息來源
Some features of POWER6 were revealed at NERSC user group meeting (http://www.nersc.gov/about/NUG/meeting_info/Jun04/). According to page 28 of this presentation

(http://www.nersc.gov/about/NUG/meeting_info/Jun04/kramer.ppt), the specifications of POWER6 will be;

peak performance: 20 Gflops,
memory bandwidth: 3.75 bytes/flop,
hardware support for ViVA(Virtual Vector Architecture)-2.

If POWER6 will have two FPU and will perform 4 flops per cycle, the frequency will be 5GHz and the memory bandwidth will be 75 GB/s.

Previously only a few details (http://www.cs.utexas.edu/users/cart/arch/fall03/KallaSlides.pdf) were known about POWER6.

The architecture will be based on POWER4/5,
It will be fabricated with 65nm process,
It will have "ultra high frequency cores".

Now, we have known the meaning of "ultra high frequency core". IBM will move to another "netburst" architecture?


如果單顆Power6的理論值是20G Flops,可能會有下面三種假設:

˙單核心,兩組FMA,5GHz x 2 x 2 = 20G Flops
˙雙核心,兩組FMA,5GHz x 2 x 2 = 20G Flops
˙雙核心,四組FMA,2.5GHz x 2 x 2 x 2 = 20G Flops,不過時脈只有2.5GHz的話,也不是什麼「Ultra High Frequency」了。

當然,也許還有更誇張的可能:20G Flops其實只是單核心的效能,那這樣Montecito大概真的就被打趴到永世不得翻身這樣。

不過,姑且不論這些可能性,每個flop可以分到3.75 bytes的記憶體頻寬,已經堪稱是超級電腦的等級(就「一般的標準」而言,大概是每個flop分到4 bytes),很好奇IBM到底在想什麼,他們都已經有BlueGene了。

5GHz加上75GB/s的理論記憶體頻寬,會產生何種結果?真是令人期待。

3 則留言:

匿名 提到...

Waterball先生:
請問IBM這樣Ultra-High Clock速求為何?

molesterwaterball 提到...

大概認為它們的設計能力和製程足以「逆向操作」吧,或著是回歸原點,採用大量簡單的核心、卻又具備高時脈的優勢等等。

匿名 提到...

請問哪裡有資料介紹In-order與OOO CPU的差別

為何很多人認為X360的Tri-core 3.2G= Dual G5,而且為何不用OOO的3.2G Dual G5

PS:台灣有圖書館放這種介紹微處理器的雜誌.實在是太貴了買不起