Chiphell - 分享与交流用户体验

查看: 4625|回复: 49

[通讯科技] 美国Summit超级计算机:采用超过27000块NV计算卡

发表于 2018-6-14 08:01 | 显示全部楼层 |阅读模式
提示: 作者被禁止或删除 内容自动屏蔽
发表于 2018-6-14 08:26 | 显示全部楼层
发表于 2018-6-14 08:35 来自手机 | 显示全部楼层
发表于 2018-6-14 08:48 | 显示全部楼层
本帖最后由 tryqandy 于 2018-6-14 08:52 编辑
stevewang 发表于 2018-6-14 08:26

Summit造价2.2亿美元 ,太湖之光造价18亿人民币
Summit功耗13MW ,太湖之光功耗15.3MW

发表于 2018-6-14 08:49 | 显示全部楼层
fairness 发表于 2018-6-14 08:35

发表于 2018-6-14 08:59 来自手机 | 显示全部楼层
tryqandy 发表于 2018-6-14 08:48
Summit造价2.2亿美元 ,太湖之光造价18亿人民币
Summit功耗13MW ,太湖之光功耗15.3MW

发表于 2018-6-14 09:02 | 显示全部楼层
tryqandy 发表于 2018-6-14 08:48
Summit造价2.2亿美元 ,太湖之光造价18亿人民币
Summit功耗13MW ,太湖之光功耗15.3MW

这不科学,光27000块NVIDIA Tesla V100计算卡价格都不至2.2亿美元,还不包括其他IBM Power处理器、连接设备等等。

  1. NVIDIA表示将会率先更新用于深度学习超算DGX-1上,内部同样集成了8张Tesla V100,提供960 TFLOPS Tensor深度计算性能,能够在8小时完成TITAN X需时8天的计算量,极大地节约科研人员等待深度计算结果时间。就当然啦,拥有8张真·核弹的新DGX-1售价并不便宜,149000美元,折合人民币102万元。
发表于 2018-6-14 09:04 | 显示全部楼层
本帖最后由 聚沙成塔 于 2018-6-14 09:06 编辑
tryqandy 发表于 2018-6-14 08:48
Summit造价2.2亿美元 ,太湖之光造价18亿人民币
Summit功耗13MW ,太湖之光功耗15.3MW

发表于 2018-6-14 09:22 | 显示全部楼层
发表于 2018-6-14 09:26 | 显示全部楼层
lhasa 发表于 2018-6-14 09:02
这不科学,光27000块NVIDIA Tesla V100计算卡价格都不至2.2亿美元,还不包括其他IBM Power处理器、连接设 ...

发表于 2018-6-14 09:32 | 显示全部楼层
本帖最后由 reaper 于 2018-6-14 09:34 编辑




发表于 2018-6-14 09:36 | 显示全部楼层

原帖地址: ... 00-petaflops-Summit

问题:Can the USA overtake China in the supercomputer race with her 200 petaflops Summit?

回答:Regrettably, no. The USA needs to aim for at least 2 exaflops if they are starting now, since it takes about two or three years to go from design to actual running hardware.

Sitting in Singapore, I am in the unusual position of having visited the Chinese #1 and #2 supercomputer sites and met with their designers, and also being familiar with the usual TOP500 sites in the US and what some of their plans are.

As a citizen of the USA, I find the disparity discouraging. The Chinese already have a 1 exaflops system in their pipeline, and their designs are excellent and getting better. By the time the USA ekes out a 200 petaflops system, the Chinese will almost certainly have one already in operation that runs circles around it.
While the USA economy grew slowly from 2001 to 2016, China’s economy grew very quickly and in some specific areas they actually surpassed the USA, supercomputing being one of them. Their supercomputer engineers did a lot of things right:

    They were willing to spend the large amount of money required to compete in this arena.
    They were not wedded to the x86 processor mentality, and instead created a hardware-software stack (mostly based on the Alpha processor and its associated software that they had licensed) that delivers significantly more speed per watt of power. Alpha processors achieve about twice as many instructions per clock as do x86 processors.
    They took a holistic approach to the physical design, from the building structure to the cooling system to the custom cabinet design, instead of restricting themselves to easily-available and low-cost solutions to those challenges.
    They bought or licensed USA technologies and picked out the best ideas, then combined them with shrewd ideas of their own.

I should probably note that my colleague Jack Dongarra has written that Taihu Light is not based on the Alpha processor. With all due respect, he should ask one of his Mandarin-speaking graduate students to read the Taihu Light literature that has not been translated into English. The use of Alpha appears in the original, but not in the English translation… I am surrounded by bilingual Singaporeans who can confirm this! There may have been sensitivity that the designers would be accused of ripping off USA technology, so they de-emphasize that aspect of their design. They shouldn’t. There was easily enough innovation in the Taihu Light to earn the designers not just one, but two Gordon Bell Prizes.
我这里应该提到我的同事Jack Dongarra写过太湖之光不是基于Alpha处理器的。说实话,他应该问问他手下的大陆研究生去读一下那些没被翻成英语的太湖之光的文章。中文版是有说用了Alpha的,但是英文翻译就没了。我周围的懂双语的新加坡人确认了此事!设计师可能是害怕会被人指责说是偷窃了美国技术,所以就没有提到设计的这方面。他们这么做很不应该。太湖之光的创新足够让设计师拿到不止一个,两个Gordon Bell奖。
Each processor of Taihu Light looks like the Cray T3D on a chip. The Cray T3D was a nimble system based on Alpha processors that many HPC people feel was one of the best-designed supercomputers of all time.
太湖之光的每个处理器都像是一个芯片上集成了一整台Cray T3D。Cray T3D是一个很牛逼的机器,基于Alpha处理器,很多超算人认为这是有史以来设计的最好的超算之一。
Most supercomputers are severely communication-bound; the T3D was much less so, with an unusually good system balance and low-latency interconnect that made it easier to sustain a high fraction of the peak rated speed. Imagine a 256-processor T3D on a single chip (together with four processors that service that array), and a cleverly-cooled cabinet that packs hundreds of those close together, and a roomful of those cabinets, and you have a system that makes the DOE and NASA labs in the USA go… *gulp*.
If the USA wants to really get back in the game and not just play catch-up, they need to break the me-too paradigm of filling standard racks with x86 processors that have GPU accelerators attached. We can get an order of magnitude improvement in operations per joule by rethinking everything. If I were doing it, I’d use a Very-Long Instruction Word (VLIW) processor with no caches, no instruction lookahead or speculative execution or branch prediction, explore the use of gallium nitride with 16-level logic instead of silicon CMOS, change the numeric representation from IEEE floats to posit arithmetic, connect the cabinets with free-space optics at terabytes per second per channel and a full crossbar, use only stacked memory and extensive use of in-processor RAM and ROM, and declare it a “moon shot” to make such a system work by 2022. This is the way all the great breakthroughs in supercomputing have been made historically… by being willing to change paradigms. Right now, the Chinese are proving better at breaking from legacy thinking than the USA.
发表于 2018-6-14 09:38 | 显示全部楼层
发表于 2018-6-14 09:39 | 显示全部楼层
本帖最后由 reaper 于 2018-6-14 09:46 编辑
lhasa 发表于 2018-6-14 09:02
这不科学,光27000块NVIDIA Tesla V100计算卡价格都不至2.2亿美元,还不包括其他IBM Power处理器、连接设 ...



这台服务器有4608个节点,每个节点拥有两枚Power9,6张NV V100,总内存是10PB.....豪就是豪,堆起硬件都透露着一股豪气.

发表于 2018-6-14 09:49 | 显示全部楼层
发表于 2018-6-14 09:49 | 显示全部楼层
lhasa 发表于 2018-6-14 09:02
这不科学,光27000块NVIDIA Tesla V100计算卡价格都不至2.2亿美元,还不包括其他IBM Power处理器、连接设 ...

发表于 2018-6-14 10:00 | 显示全部楼层
reaper 发表于 2018-6-14 09:39
我研究了一下这个项目,其实是能源部牵头,有IBM,NV,田纳西橡树岭国家实验室(没错,就是美国nuke的发源地)联 ...

发表于 2018-6-14 10:03 | 显示全部楼层
发表于 2018-6-14 10:03 | 显示全部楼层
wenzi2203 发表于 2018-6-14 09:26

发表于 2018-6-14 10:07 | 显示全部楼层
Filter 发表于 2018-6-14 09:49
不管怎么说,老黄还真是挺厉害的,各个提前布局的领域现在都开始领先业界了,果然格局和眼光不同凡响呀! ...

发表于 2018-6-14 10:32 | 显示全部楼层
"如果我来干这事情的话,我会用一个超长指令处理器(VLIW),不带缓存,不要向前检测,推测执行或者分支预测,将数字表示从IEEE浮点换成别的模式" -- 这说的不就是GPU嘛
发表于 2018-6-14 10:41 | 显示全部楼层
reaper 发表于 2018-6-14 09:36

原帖地址:https://www.qu ...

发表于 2018-6-14 10:48 | 显示全部楼层
lhasa 发表于 2018-6-14 09:02
这不科学,光27000块NVIDIA Tesla V100计算卡价格都不至2.2亿美元,还不包括其他IBM Power处理器、连接设 ...


发表于 2018-6-14 10:51 | 显示全部楼层
farwish 发表于 2018-6-14 10:41
感谢翻译,,,美国人就是实事求是,,自己不足就是不足,从不会诋毁别人吹嘘自己,我觉得这才是一个民族 ...

发表于 2018-6-14 11:16 | 显示全部楼层
本帖最后由 farwish 于 2018-6-14 11:21 编辑
reaper 发表于 2018-6-14 10:51


不可否认自夸这点 某国第二没人敢说第一吧

再说有色眼镜,觉得在中国政府带节奏下,,现在中国人对美国的有色眼镜远大于美国队中国吧  只是一个是政治上  一个是经济上
发表于 2018-6-14 12:06 | 显示全部楼层
farwish 发表于 2018-6-14 11:16

不可否认自夸 ...

不对  是美国人都给洗脑洗傻了
发表于 2018-6-14 12:11 | 显示全部楼层
tryqandy 发表于 2018-6-14 08:48
Summit造价2.2亿美元 ,太湖之光造价18亿人民币
Summit功耗13MW ,太湖之光功耗15.3MW

发表于 2018-6-14 12:15 | 显示全部楼层
farwish 发表于 2018-6-14 11:16

不可否认自夸 ...

发表于 2018-6-14 12:25 | 显示全部楼层
catabao 发表于 2018-6-14 12:15
我擦...论自夸某国离美国差远了好吧...你是真没接触过美国人还是怎么地。。。 ...

意淫强国也不知道说的某国,,说自夸你是没看过全美 指责政府,抨击总统,亡国论的书 可以随意流行

接触倒是接触过 呆过四五年吧

感觉他们还是满谦虚的,就事论事,当然上升到两国之间 意识形态差异,肯定是各看不惯各
发表于 2018-6-14 12:26 | 显示全部楼层
emu10kx 发表于 2018-6-14 12:06
不对  是美国人都给洗脑洗傻了

嗯 这倒是,,两国意识形态差异,都在放渗透,各看不惯各 倒是有的
您需要登录后才可以回帖 登录 | 加入我们


小黑屋|手机版|Archiver|Chiphell ( 沪ICP备12027953号-5 )沪公网备310112100042806

GMT+8, 2018-12-17 01:36 , Processed in 0.035412 second(s), 13 queries , Gzip On, MemCache On.

Powered by Discuz! X3.1

© 2007-2017 All rights reserved.

快速回复 返回顶部 返回列表