POPPUR爱换

标题: Xeon Phi 协处理器正式发布，Larrabee->MIC->Xeon Phi [打印本页]

作者: Edison 时间: 2012-6-19 08:24
标题: Xeon Phi 协处理器正式发布，Larrabee->MIC->Xeon Phi
Today at the International Supercomputing Conference (ISC) 2012 in Hamburg, Germany, Intel announced their Knights Corner coprocessor, the first member of their XEON Phi family. Knights Corner is the next generation of their multi-core, parallel processing coprocessor for the HPC (High Performance Computing) race to be a contender in the TOP500 Supercomputing Sites.

[attach]1920035[/attach]

Knights Corner specifications

[attach]1920037[/attach]

1 Teraflops from coprocessor node of Intel's HPC cluster

FLOPs (or flops or flop/s, for floating-point operations per second ) are a measurement of a computer's performance (e.g., required by a given algorithm or computer program). Way back when scientists starting counting FLOPS, human nature obviously sprung into action and the race was on, for the “mine is faster than yours” TOP500 Supercomputing fastest FLOPS trophy. The race escalated to teraflops which are computing speed equal to one -million million-floating-point operations per second. Then came petaflops at quadrillion (thousand trillion) floating point operations per second (a thousand teraflops or 2 to the 50th power FLOPS).

The TOP500 project ranks and details the 500 (non-distributed) most powerful computer systems known in the world. This is the most important ranking in the HPC world of applying advanced computational science and engineering to critical fields like renewable energy, genomics and biosciences, nuclear safety, weather studies, design and simulation to name a very few. Underlying the fastest FLOPS race is the fact players in the TOP500 rankings also make lots and lots of profit from their highly sophisticated reuse of existing technology.

[attach]1920036[/attach]
From FLOPS to Petaflops

Rajeeb Hazra, Vice-President Intel Architecture Group & GM Technical Computing, said last November they demonstrated the first silicon of the Intel Xeon Phi coprocessor, code named “Knights Corner”. It produced an astounding teraflop of performance in a processor the size of your thumb, setting the industry on notice of the potential of many core architectures and providing a clear path of how we'll get to the Petascale and Exascale era. This is the same amount of flops as the Number one supercomputer in 1997on the TOP500 list, dubbed ASCI Red. ASIC Red used thousands of processors and filled a room with cabinets to produce the same amount of performance. Knights Corner quickly got the nickname of “Supercomputer on a Chip”.

Last year at ISC 2011, Intel announced MIC (Many Integrated Core) architecture. MIC was not an outstanding branding of their engineering ideas. So earlier this year, Intel decided to improve their Fabric Technology which provides the communications links for data flow between processors and I/O devices.

In January, they bought InfiniBand from Qlogic for $125 million. This gave them scalable technology which is used to connect servers in high-performance computing (HPC) environments. Then in April, Intel made a very brief announcement that they were acquiring Cray's Aires Technology for their HPC programs. Diane Bryant, vice president and general manager of Intel's Datacenter and Connected System Group said: "The acquisition of Cray's industry-leading interconnect technology and expertise provides exceptional strategic assets that further enhance Intel's HPC portfolio. We're excited about the value this will allow us to bring to our customers."

This is the same Ms. Bryant who announced this spring the XEON E5/2600 server line. When asked about Intel losing out to AMD over Seamciro's architecture and fabric Ms. Bryant said Intel "did look at SeaMicro's fabric technology. There are probably very few people they didn't come to and shop their solution to. We were not impressed. We declined and very soon after our competitor acquired them." There does seem to be a contradiction in some of Intel's explanations about HPC, server fabric technology, and whose stuff works - moving right along.

The target market for XEON Phi and Knights Corner are big universities and National Labs who are existing Intel customers with HPC-style, X86-based applications.

[attach]1920038[/attach]

Target market for XEON Phi family

Those folks have old X86 apps and they were interested in XEON Phi and Knights Corner because they won't have to re-write their old stuff into GPGPU (general purpose Graphic Processor Unit) languages like Nvidia's CUDA or open-source OpenGL. Probably the decision by Oak Ridge National Labs to not rewrite their X86 apps were driven more by budget justification problems than anything else. In other words, if your X86 coding ain't broke, don't rewrite it for a GPGPU, simply add a bigger "engine" like a Knights Corner node to your HPC cluster.

In our next article about Knights Corner and the XEON Phi family, we will investigate where they came from and what it really means in light of Intel's successes in this year's TOP500 HPC FLOPS race.

作者: nBAN 时间: 2012-6-19 08:31
最终还是用的 pcie 目前top500有一台排在150位这次top500 ibm蓝色基因q狂胜红杉（16pflops）第一米拉（8pflops）第三

作者: cool_exorcist 时间: 2012-6-19 08:46
头4名全是纯cpu

作者: vironlee 时间: 2012-6-19 09:47
不识货的以为是Intel独立显卡……

作者: Edison 时间: 2012-6-19 11:06
大图一枚：
[attach]1920181[/attach]

作者: nBAN 时间: 2012-6-19 11:33

cool_exorcist 发表于 2012-6-19 08:46
头4名全是纯cpu

其中3个就是ibm的

作者: nBAN 时间: 2012-6-19 11:36

Edison 发表于 2012-6-19 11:06
大图一枚：

总觉得这单颗1tflops的协处理器要组成1eflops的超级计算机是不是使用机柜太多？占地面积太大？

作者: elisha 时间: 2012-6-19 11:39
1T双精度，恐怖了

作者: mooncocoon 时间: 2012-6-19 12:02
Intel的野望竖起来了，有ISA，有编译器，有cash，有突破物理限制的能力和需求，剩下的就看它有没有彻底的决心用这个比CPU更方便并行的方式玩HPC了。

作者: asdfjkl 时间: 2012-6-19 12:03
这个牛掰的。。。
给NV和AMD的时间不多了~~~

作者: zllywx 时间: 2012-6-19 12:11
LZ的头像蛮有爱的。。

作者: NORAWITHMYCALL 时间: 2012-6-19 12:32
拉拉比的遗产，砍去图形输出功能，全力提高浮点运算，接下来，大家等着看好戏吧，NV的CUDA GPGPU，AMD APU&AMD CPU+GPGPU，INTEL CPU+MIC，IBM等的纯CPU，看谁能笑傲HPC领域。

作者: nBAN 时间: 2012-6-19 12:39

NORAWITHMYCALL 发表于 2012-6-19 12:32
拉拉比的遗产，砍去图形输出功能，全力提高浮点运算，接下来，大家等着看好戏吧，NV的CUDA GPGPU，AMD APU& ...

看看今年的top500 我相信最有竞争力的还是ibm 其次是intel

作者: kinno 时间: 2012-6-19 12:40
cuda就一浮云

作者: elisha 时间: 2012-6-19 13:13

nBAN 发表于 2012-6-19 12:39
看看今年的top500 我相信最有竞争力的还是ibm 其次是intel

又不冲突，ibm是做系统的，intel是做芯片的，ibm的系统大部分也是用intel的芯片

作者: asdfjkl 时间: 2012-6-19 13:22

kinno 发表于 2012-6-19 12:40
cuda就一浮云

搞笑了。。。。。 cuda是08年发布的，Intel当时还在造显卡在。
软件每家都得花大力气，而且这方面Intel是高手，AMD的策略是节省财力，靠第三方，搞联盟，那是真真的死翘翘。

作者: NORAWITHMYCALL 时间: 2012-6-19 13:32

nBAN 发表于 2012-6-19 12:39
看看今年的top500 我相信最有竞争力的还是ibm 其次是intel

基本同意，IBM目前还毋容置疑的强大，坐稳第一把交椅还是最有可能的。其实个人比较关心，AMD的APU组成的HPC和GCN架构的FIRESTEARM作为GPGPU在HPC市场中的作为，如果这2个成功了，以后NV的好日子将会变得不怎么好过。

作者: kinno 时间: 2012-6-19 13:34

asdfjkl 发表于 2012-6-19 13:22
搞笑了。。。。。 cuda是08年发布的，Intel当时还在造显卡在。
软件每家都得花大力气，而且这方面Intel是 ...

别说我是N黑，记得当年cuda出来时，很多N粉叫喧gpu会取代cpu的
有时候，联盟体制并不适合某些事务的推广，相反，intel自成一家倒是简介流程，人多不一定能解决复杂问题

作者: NORAWITHMYCALL 时间: 2012-6-19 13:35

elisha 发表于 2012-6-19 13:13
又不冲突，ibm是做系统的，intel是做芯片的，ibm的系统大部分也是用intel的芯片

恰恰是IBM是做打包整体解决方案的，从硬件到软件一并卖给你，也就是POWER7服务器+AIX全盘兜售，这才是IBM实力的真正体现。

作者: disruptor 时间: 2012-6-19 13:51
22nm终于把功耗压下来了

作者: zaknafein 时间: 2012-6-19 13:58
注意这个东西不需要cpu也能跑系统的，准确得说，自己就是跑嵌入式linux，外面都是走客机访问模式

作者: nBAN 时间: 2012-6-19 14:04

elisha 发表于 2012-6-19 13:13
又不冲突，ibm是做系统的，intel是做芯片的，ibm的系统大部分也是用intel的芯片

错了吧 ibm的x86解决方案从来都最底层的蓝色基因的芯片都是ibm自家的这也是ibm牛b的原因

作者: nBAN 时间: 2012-6-19 14:05

zaknafein 发表于 2012-6-19 13:58
注意这个东西不需要cpu也能跑系统的，准确得说，自己就是跑嵌入式linux，外面都是走客机访问模式

能吗？这是协处理器

作者: luckissy 时间: 2012-6-19 14:42
IBM是提供完整的方案哈哈蓝色基因Q的能耗比很牛B了

作者: zaknafein 时间: 2012-6-19 14:48

nBAN 发表于 2012-6-19 14:05
能吗？这是协处理器

能，叫协处理器是为了能和xeon挂钩卖，市场部门的决定

作者: badchris 时间: 2012-6-19 15:07
提示: 作者被禁止或删除内容自动屏蔽

作者: disruptor 时间: 2012-6-19 16:42
2018年的一个刀片竟然能有0.5-1p。nv的echelon不是说单芯片也只有20t左右的性能吗。intel这是什么科技，6年能增加500倍的性能？

作者: amuly 时间: 2012-6-19 16:57
不知道这东西能不能在hpc这边站住脚跟以后，intel继续搞搞larrabee

作者: badchris 时间: 2012-6-19 17:09
提示: 作者被禁止或删除内容自动屏蔽

作者: badchris 时间: 2012-6-19 17:10
提示: 作者被禁止或删除内容自动屏蔽

作者: amuly 时间: 2012-6-19 18:31

badchris 发表于 2012-6-19 17:10
http://semiaccurate.com/2012/06/12/what-comes-after-knights-landing/
Larrabee应该是以色列那边的人 ...

这玩意跟搞cpu不一样了吧，

要搞显卡，物理设计一方面，驱动分量也不轻啊

先解决了成千上万的游戏的兼容问题，不至于出现贴图错误，跳出这些

然后至少要把流行的有活跃人数的几百上千个游戏的性能提到可以接受的地步

我感觉不是一个以色列的团队在默默研究能搞定的事情

作者: gz_easy 时间: 2012-6-19 18:35
"Coprocessor"

好熟悉的名字， NNNN年前，Intel就是这样称呼它的CPU协处理器

作者: elisha 时间: 2012-6-19 19:45

amuly 发表于 2012-6-19 18:31
这玩意跟搞cpu不一样了吧，

要搞显卡，物理设计一方面，驱动分量也不轻啊

这个又不是显卡，哪需要考虑这些

作者: amuly 时间: 2012-6-19 20:55

elisha 发表于 2012-6-19 19:45
这个又不是显卡，哪需要考虑这些

我说的是，

经由xeon phi在hpc站稳脚跟以后

有了硬件基础和生态环境，larrabee是否会复辟

毕竟当初larrabee一个主要目的就是做显卡。

作者: elisha 时间: 2012-6-19 21:33

amuly 发表于 2012-6-19 20:55
我说的是，

经由xeon phi在hpc站稳脚跟以后

显卡市场太小，intel做集显就够了

作者: gzpony 时间: 2012-6-20 11:36

asdfjkl 发表于 2012-6-19 12:03
这个牛掰的。。。
给NV和AMD的时间不多了~~~

当年失败给NV和AMD的原型，只是换了工艺，换了个马甲就能赢？
这货不是CPU，intel要从头建立整个环境，能不能生存还是个问题，先别想那么远吧。

作者: nBAN 时间: 2012-6-20 14:10

gzpony 发表于 2012-6-20 11:36
当年失败给NV和AMD的原型，只是换了工艺，换了个马甲就能赢？
这货不是CPU，intel要从头建立整个环境，能 ...

这是协处理器原来那是显卡都不是一个概念再说这玩意比nv的特斯拉好多了

作者: gzpony 时间: 2012-6-20 14:55

nBAN 发表于 2012-6-20 14:10
这是协处理器原来那是显卡都不是一个概念再说这玩意比nv的特斯拉好多了

是同一个概念，就像NV的geforce和tesla一样，只是用途的定位不同。

这东西做显卡的时候连上市挑战geforce的勇气都没有，现在为什么就有人敢说比nv的特斯拉好多？
有什么实际的测试和数据支持？

作者: badchris 时间: 2012-6-20 14:58
提示: 作者被禁止或删除内容自动屏蔽

作者: elisha 时间: 2012-6-27 22:53

gzpony 发表于 2012-6-20 14:55
是同一个概念，就像NV的geforce和tesla一样，只是用途的定位不同。

这东西做显卡的时候连上市挑战gefo ...

术业有专攻，这货是专精计算的

作者: potomac 时间: 2012-6-27 23:52
提示: 作者被禁止或删除内容自动屏蔽

欢迎光临 POPPUR爱换 (https://we.poppur.com/)