Deep learning host save machine notes

Update: This article was written a year ago. This year , the tide of deep learning continued to advance. The 1080 has also been upgraded to the 1080TI, and there are more and better options for builders. Recently updated an article: " Building a Deep Learning Server from Scratch: Hardware Selection ", you can read the following (mainly provide some ideas for selection), and then look at the latest one (mainly provide some configuration options), Complement each other. In addition, the host of Thunder Century is strongly not recommended, and the after-sales service is seriously unreliable.

Several articles have been written in this series, here is an index of related articles for reference only:

In mid-to-late May 2016, the announcement and sale of GTX1080 directly stimulated my desire to build a deep learning host. For me, it has been more than ten years since the first PC was saved in college. , In fact, within the budget of 5,000 yuan, I went to the computer city to find a business to assemble a desktop computer, which is called DIY.

Although the graphics card has been locked, it is still very vague for other collocations, just need "good CPU", "big memory", "big hard disk", so I started to google "deep learning computer", "deep learning server", "deep learning PC" ", "Deep Learning Host", "Deep Learning Machine", "Deep Learning Workstation" These keywords, and quickly locked this article " How to Build a Deep Learning Server " as the main reference:

Hardware selection: The basic idea is a single graphics card machine, leaving room for upgrades

......

CPU selection:
In deep learning tasks, the CPU is not responsible for the main task. When a single graphics card is calculated, only one core reaches 100% load, so the number of cores of the CPU and the number of graphics cards can be the same, too many are unnecessary, but the bandwidth of processing PCIE to 40.

Motherboard selection:
It needs to support X99 architecture, PCIe3.0, and 4-channel DDR4 memory architecture. If you want to engage in four graphics cards in parallel, the PCIE bandwidth support must reach 40, and support 4-Way NVIDA SLI technology.

Memory:
double the video memory, of course, if you have money, the bigger the better.

Power supply problem: The power of one graphics card is close to 300W, and the recommended power supply of four graphics cards is more than 1500W. For future expansion, a 1600W power supply is selected.

Chassis heat dissipation:
Because the various components are quite large, a large chassis with good heat dissipation function is required, and the Tt Thermaltake Core V51 chassis is selected, with three 12cm fans as standard. In the future, if necessary, water cooling equipment can be installed.

......

Final hardware configuration:
CPU: Intel X99 platform i7 5960K
Memory: DDR4 2800 32G(8G*4)
Motherboard: GIGABYTE X99-UD4
Graphics Card: GTX Titan X
Hard Disk: SSD+Ordinary Hard Disk

按照这篇文章的配置,在淘宝查了一下相关价格,这个配置大概20000+的水平,如果作者再搞个4路显卡并行,绝对土豪。不过这里面的i7 5960K很少见,应该对应的是i7 5960X。

对于我来说,主要是围绕着GTX1080来配置深度学习服务器,不过这个时候上文所说的一些概念还比较模糊,于是以一个游戏玩家的身份跑了一趟电脑城,当我向商家说明来意,要配置一个GTX1080主机时,商家除了纷纷表示GTX1080货源紧缺,需要预定外,也有的会在打完几个电话后告诉我有现货但需要加价,极端的一个例子是华硕的公版GTX1080当时加到了6900。不过几乎无一例外,他们给的配置清单基本上是这样的:

CPU: Intel i7 6700K
内存: DDR4 32G(8G*4)
主板: 华硕Z170-P or Z170-AR
显卡: GTX 1080
硬盘: SSD+普通硬盘
......

加上其他配件,大概12000+的样子。当然,我只是通过市场调研一下,并不急于入手,于是回到家里,继续google。

i7-6700K 是去年发布的6代酷睿i系列处理器,它最大的变化是 “14nm工艺+新架构(Skylake),对于老态龙钟的Haswell有着工艺和架构的双'料'改进!新架构意味着同频性能更强,新工艺意味着功耗更低”。另外i7-6700K是“四核心八线程、8MB三级缓存,CPU频率基准4.0GHz、最高4.2GHz,总线支持十六条PCI-E 3.0,内存支持双通道DDR4-2133、DDR3L-1600”。特别注意这里的PCIE 3.0,总计只有16条,这个就很有局限了。同样它所支持Z170等系列主板,无论在显卡支持数量和最大内存容量上都不如X99的可扩展性强。譬如这里的华硕Z170主板,必须是Z170-A或者在Z170-AR才能支持双路显卡,稍差一点的Z170-P仅支持一个显卡。内存插槽上,Z170多是4个,而X99一般都是8个。

于是将目光又一次转向了X99平台,重新审视了i7-5960Xi7-5930Ki7-5820k这几个两年前英特尔发布的22纳米工艺的Haswell-E系列的CPU。下面这幅图其实一目了然:

5960x5930x5820

注意其中的PCIE-3.0个数,5960x和5930k都是40,而5820K只有28,这样的话“具备完整的40条PCI-E的i7-5960X和i7-5930K可以以“x16+x8+x8+x8”的带宽分配方式组建四路的SLI或者CrossFire,而“小弟”i7-5820K最多只能组建“x16+x8”或“x8+x8+x8”的双路或者三路显卡并联系统。”

显然,在可扩展性上,5960x和5930k更好,不过在价格上自然也有区分,目前淘宝上5960x大概7000左右,5930k在4000左右,5820k在2700左右。

对于我来说,还没有打算做多路显卡并联,但是为了可扩展性,所以初步选择了5390k或者5820K作为CPU备选。但是最大的问题还是GTX1080的货源问题,淘宝天猫上全是预定,托朋友关系咨询当地技嘉总代,拿GTX1080必须搭配其他显卡销售,不过有个很优惠的价格可以拿到 GTX Titan X ,为此还在微博上咨询了一下是选择GTX1080还是GTX Titan X,李沐M和其他几个同学的回答基本让我确定等待GTX1080。另外李沐大神有几篇关于GPU的文章,不过都是超级土豪的,感兴趣的同学可以参考。

之前曾google到一篇GTX1080机器的文章:国产首款GTX 1080游戏PC开售,不过当时对这些配置还无感,调研了一番回头再看的时候,发现雷霆世纪所推的这款GTX1080主机性价比超高,可惜这个主机第一轮预定完之后到目前为止一直显示无货:

CPU:Intel i7-6800K(6核12线程)
主板:华硕X99-E
显卡:GTX 1080
SSD:三星SM951 256G(M.2接口)
电源:海韵X-850 850W(80 PLUS全模组)
机箱:迎广805C红 中塔式机箱(铝合金,双面钢化玻璃侧透)
散热:采融B81 V2黑色雷霆定制版(纯铜底、6热管、PMW风扇)
内存:自行购买(DDR4)
系统:预装Windows 10测试版系统

这套配置里面提到了i7-6800K,查了一下,是最近Intel刚刚推出的发烧级桌面处理器Broadwell-E系列中的一员:

6800k

网上有评论i7-6800k是i7-5820k的升级版,虽然同样支持X99架构主板,但是同样的问题也是PCIE-3.0个数只有28个。不过同为升级版,不差钱的同学可以考虑将上文中i7-5960X的配置替换为Broadell-E系列里的旗舰产品i7-6950X,10核心20线程顶级配置,淘宝报价目前在15000左右。

雷霆世纪这款GTX1080主机虽然很诱惑,但是一直显示无货,不过在浏览相关的产品的时候发现另一款 “The one 2 Plus自由版” ,除了没有显卡,主板型号略微有点差异外,其他配置和上一款GTX1080机器基本相同,当然都没有内存。此时恰逢京东618期间的活动,价格比原价7488还低了600,6888可以搞定,稍微淘宝了一下相关的配件价格,粗略计算自己配的话大概需要8500+,所以马上付定金预定。当时的页面显示,6.26号付全款,6.28号之后按付款顺序发货。

所以必须等了,不过这期间一个朋友通过种种努力在当地技嘉总代帮我搞定了一块非公版GTX1080显卡:技嘉GTX1080 G1 GAMING ,于是,在拿到显卡的当天,发了一条微博:为信仰充值。之后又在淘宝上买来了4条16G内存条和一块4T硬盘(组SSD+普通硬盘),外加显示器和鼠标硬盘,这套所谓的GTX1080深度学习主机配置是这样的:

CPU:英特尔(Intel)酷睿六核i7-6800K 2011-V3接口 盒装CPU处理器
主板:华硕(ASUS)X99-A/USB 3.1 主板 (Intel X99/LGA 2011-v3)
显卡:技嘉GTX1080 G1 GAMING 非公版
硬盘:三星 SM951 M.2 256G SSD + 西部数据 WD40EZRZ 4T台式机硬盘(蓝盘64M)
内存:64G 金士顿骇客Fury DDR4 2400 16G单条 * 4
机箱:迎广(IN WIN)805c红 ATX中塔式机箱 黑紅色 铝合金/钢化玻璃/双面侧透(U2*2+U3*1+U3.1*1)
电源:海韵(Seasonic)额定850W X-850 电源(80PLUS金牌/全模组/全日系电容/支持SLI/支持背线)
散热器:采融 B81 V2(黑色)

总计15000多一点:主要配置(6888)+ 显卡(5000)+ 4条内存(1800,最近内存涨得比较猛)+ 硬盘(780)+ 显示器(800)。前几天终于拿到主机,并找来雷霆售后师傅帮我把散热、内存、显卡、硬盘安装调试好,系统预装的是Windows10试用版,当然,很快它就会被Ubuntu取代,最后上几张图:

296275349

614287776

1743460326

274517047

参考资料:

如何搭建一台深度学习服务器:http://www.r-bloggers.com/lang/chinese/2042
GPU集群折腾手记——2015:http://mli.github.io/gpu/2016/01/17/build-gpu-clusters/
Nvidia新的Pascal值不值得买(升级):http://mli.github.io/2016/06/14/new-pascal/
如何配置一台适用于深度学习的工作站:https://www.zhihu.com/question/33996159
Which GPU(s) to Get for Deep Learning: http://timdettmers.com/2014/08/14/which-gpu-for-deep-learning/
A Full Hardware Guide to Deep Learning:http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/
Building a Deep Learning (Dream) Machine:http://graphific.github.io/posts/building-a-deep-learning-dream-machine/
Reddit: gtx1080 vs 1070 for machine learning?

注:原创文章,转载请注明出处及保留链接“我爱自然语言处理”:http://www.52nlp.cn

本文链接地址:深度学习主机攒机小记 http://www.52nlp.cn/?p=9081

Deep Learning Specialization on Coursera
This entry was posted on July 5th, 2016 . It belongs to the deep learning category and is tagged with  GTX1080 , GTX1080 host , host , workstation , Gigabyte GTX1080 , server , deep learning , deep learning PC , deep learning host , deep learning workstation , deep learning server , deep learning machine , deep learning computer  . The author is 52nlp .

Guess you like

Origin http://43.154.161.224:23101/article/api/json?id=325182628&siteId=291194637