800 yuan eight core server? second hand server building guide

Posted by barello at 2020-04-16

*Original author: push, eax, this article belongs to freebuf original award program, reprint is prohibited without permission

When you spend nearly ten thousand yuan to chop i7 5960x, have you ever thought that in a corner of Huaqiang North, a group of people rely on hundreds of second-hand server accessories to build a server with the same or even stronger performance!

I don't know how many people still remember the post "see how I can build a workstation with eight cores and 16 threads and 32g memory with 1000 yuan". The author randomly sent (Dian) a (GAO) post (Fei), did not expect to get nearly 100 replies! After all, it's in this field. I think you must be interested in hardware.

In the last post, the author showed the whole process of building a second-hand server by himself. Many students were confused, so the author had the intention to post this post.

In this post, I will try my best to systematically describe the garbage collection methods and accessories worth cutting hands, hoping to inspire you and build your favorite server.

First of all, before reading this post, please make sure that you have the experience of saving machines and a heart that can stand the twists and turns (or that you have a lot of money).

The process of saving foreign garbage server is similar to that of ordinary computer (in general), you only need to have your own experience of installation.

Long text, multi map warning, but it is recommended to read the whole article carefully, which is the author's many years of experience.  

Next, we will continue to write the second-hand server (workstation) building guide.


1. Garbage collection overview

2. Server CPU overview

3. Accessories (memory, hard disk) Introduction

4. Overview of main board (including quasi system)

5. Lga1366 details

6. Lga2011 (including V1 and V3)

7. Introduction to my common configuration list

8. 800 yuan 8-core construction practice

Because the full text is too long, it is divided into two parts: the first three chapters are the top, focusing on concepts and basic knowledge; the last five chapters are the bottom, focusing on practice.


1. Most of the information in this article is from the author's experience and collected everywhere. It is inevitable that there are mistakes and omissions. Please do not hesitate to comment.

2. Some content and pictures of the article are extracted from the Internet. If there is any infringement, please contact the author.

3. All price information comes from Taobao and Huaqiangbei quotation on March 24, 2017.


brief introduction

"Garbage collection" is a nickname for tossing all kinds of second-hand server accessories. People who toss these are also known as "garbage men". Because most of the second-hand server accessories come from the evil capitalist countries, they are also known as "foreign garbage".

Foreign waste sources

These so-called foreign wastes are mainly discharged from foreign servers or OEM channels. Businesses buy them directly from developed countries such as Europe, the United States and Japan, and ship them back to China for sale (based on containers), and then sell them to retail investors by classification.

Server market accessories are often the high-end untouchable in the civil market, and the update speed is very fast, which causes a large number of second-hand accessories to be eliminated to the market. Although they have been eliminated in the server field, the skinny camel is bigger than the horse, and they can still dominate in the civil field.

Who should pick up rubbish

1. People with high requirements for hardware, especially professional needs: virtualization, open more, build stations, design, drawing .

2. People with low budget: thousands of yuan want to build a better computer, but they can't afford the 7-generation i7.

3. Fancier, like me, a computer + two servers + a balcony accessories, built not for use, but purely to meet a kind of original male desire.

Advantages of garbage collection

1. Cheap:

There is a huge amount of machine dismantling, extremely low cost, extremely asymmetric industry information, a large number of second-hand accessories, but only a few people will be tossed, resulting in the embarrassing situation of high value and low price.

Let's make a comparison:

W3520: 1366 pin server CPU i7 920: Civil core generation CPU

The two CPUs belong to the same era and the same interface, with similar parameters and performance. The only difference is that the former belongs to the CPU dedicated to the server.

It can be seen that the i7-920 can still sell at a high price of more than 400 yuan today, while the w3520 has reached the horrible price of 10 yuan cabbage!

2. Special performance optimization:

The original intention of server accessories design is to meet the most demanding requirements in the industry. Even the starting level hardware is enough to kill a large number of civil accessories.

And because of its own positioning, the server CPU is inclined to the needs of industrialization: focusing on data computing power, high stability, high reliability, high security and so on.  

The most obvious example is that many CPU servers tend to have multi-core performance. They can not move 8 cores and 10 cores, but only have more than 2 main frequencies. They play online games (only do single core optimization) to be suspended by the I3 (the I3 single core performance takes off, which is known as the I3 silent second full stop).

In fact, this design is for industrial optimization. For example, virtual machine players know that each core can be fully utilized when it is opened more than once. At this time, the most important thing is not the single core performance, but the number of cores.

(of course, there are CPUs for single core optimization, which will be mentioned later)

In addition, there are other optimizations, such as ECC memory, redundant power supply and so on. I believe that many of them are needed by all of you, and it is difficult to meet them in the civil field.

Stability problem

For second-hand parts, the most concern should be stability.

In fact, even second-hand products in the server field have the same stability as some high-end accessories in the civil field.

After entering the computer room, you should know the environment of the computer room: dust-free, constant temperature, UPS, and it is impossible to have over frequency. In a few years, the ash has declined.

Moreover, the server accessories are designed according to the most severe environment, and the materials are also the top accessories. I don't know where they are higher than the main board of any domestic work package!

For example, the number of common home motherboards is about 4500, and the number of motherboards with slightly better positioning is about 7800. For example, some motherboards with similar saber toothed tiger level can reach about 12000, and the top player's national series can reach about 1560 at most, while the ultramicro (server motherboard manufacturer) home series motherboards (Z87, etc.) In the case of six power supply, the number of main board components can reach about 1800, which can be used on the premise of ensuring the stability of the main board, without deep design power, it can not be done.

So, as long as it's not a sword, it's a second-hand server (if you play a fake board or even punch the x79 of the fly line That really depends on people's character. Stability is not a problem to worry about!

Naming rules of CPU interface

This should belong to common sense.

LGA XXXX: refers to the pin type of CPU interface. Each generation of CPU has its own independent interface type. It is customary to classify CPUs by pin.

The interface types of recent generations of server CPUs:


Server CPU naming rules

First of all, we need to know that there are three levels of server CPU, similar to core's I3 / i5 / i7. The server is named E3 / E5 / E7.

It should be noted that early server CPUs did not use "e * -" as a prefix, but use "L", "e", "X" without "-" as a prefix to represent power consumption. Followed by a number such as "x5470":

X stands for high power consumption, 5 stands for mid-range products (corresponding to the current E5); 4 stands for generation 4, that is, LGA771 interface; 7 stands for the 7th performance in this generation (the bigger the better); the last 0 suffix also stands for performance, which is occasionally used, such as x5472.

In the era of lga2011, the naming method became e * - * * * * V *, such as e5-2620l V3:

E5 represents the middle end product;

2 represents that it corresponds to a two-way server (there are also Series 1 and 4);

6. In addition to 6, 2,4,8:2 are E3, 4 are lga1356, 6 are E5, 8 are E7. Generally speaking, it is unnecessary to care about it;

2 represents that its performance in this generation is in the second place (the bigger the better);

The last 0 still has no egg tail.

But the difference is that this time the power consumption is put after the number - "L" stands for low power consumption (not necessarily with this suffix);

Algebra is at the end - "V3" stands for three generations.

Exercise: x5550; e5-2699 V4


X5550: midrange (similar to i5), generation 5 (LGA1366), fifth in performance, high power consumption

E5-2699 V4: middle end (similar to i5), E5 series generation 4 (LGA 2011 V3), highest performance ranking, standard power consumption (regardless of E7, this u is the most powerful server CPU now)

In this paper, we mainly discuss the middle end, that is, E5 CPU. In addition, there are E3 series (no cost performance) and high-end E7 series (too high-end to see).

As for the corresponding motherboard interface, according to Intel's rules, E3 and the common motherboard in the civil market have the same pin (except E3 V5); E5 is the flagship series corresponding to the civil market, such as i7 5960x and E5 2660 V3 are lga2011 V3; as for E7, it's another world. It belongs to lga1567. There is very little circulation in the market, and a main board can reach tens of thousands, which leads to no value of E7.

Workstations and servers

In this paper, the author will discuss both of them and call them servers.

Both are products of the industrial sector, but they are different:

Workstations focus on humanization, such as fast startup time, small size, and many places are similar to ordinary computers, which can be used as super high configuration personal computers.

As shown in S30 (Lenovo's 2011 workstation, I think Lenovo's design looks good):


The server is a pure industrial design, some servers for stability, a post can be as long as 3 minutes. There is also no integrated sound card, noise, volume weight (2U server can weigh up to 30kg) and other common problems.

Now I have a server equipped with six 1w5 turn violent fans beside me. It's like a helicopter. It can't be used at home without changing the speed


I personally suggest that the first time you toss a workstation, you should be familiar with the routine before you play with the server. The server is variable, some can be powered on in one minute, some can be powered on in three minutes; some can be powered off by fans, and some can't. It is recommended to do a good job of inspection before the server is tossed.

Main board and quasi system

Here we introduce two concepts: motherboard and quasi system.

Motherboard naturally needless to say, we all know what it is.

But what I want to focus on here is: because the CPU of the server series uses the most civil top-level motherboard (Intel's rule is that the top-level interface and the server are common, such as lga2011), and the amount of disassembly of the motherboard is less than that of the CPU, in many cases, the server motherboard is more expensive than the server CPU, and the water is also relatively deep, which needs attention.

So what does quasi system mean?

Most of the time, the main board of the server is strange and it's difficult to match the chassis; moreover, the power interface is not designed according to common sense, and from time to time, something like 10PIN will pop up.

For example, the shape of this motherboard is also missing:


So some businessmen simply package and sell the chassis, power supply and radiator to you. You can add CPU, memory, hard disk and video card, which is the quasi system we call.

This kind of playing method is more convenient. Like the motherboard (HP Z600 workstation) in the figure above, if you don't use the original case, you can basically choose to run naked, or build a shoebox case like this:

Let me tell you a little secret: in fact, the stability and quality of the quasi system are generally better than that of the single motherboard. The quasi system motherboard is usually a real first-hand dismantle, while the single motherboard is relatively deep. As for why you can think about it yourself.

However, the principle of "one cent, one share of goods" is always correct. Quasi system is often more expensive than parts purchased. As for which to choose, different people have different opinions. You can see your DIY ability to choose by yourself.

Where to buy foreign garbage

To buy these small accessories, e-commerce only has Taobao family to do this. The first choice for online shopping is to go to Taobao. Most of the accessories can be found.

However, in fact, I suggest physical stores, such as Huaqiangbei (I'm in Shenzhen, go directly to Baotai server to go home to Meizizi! It's that the passers-by stare at me strangely.

There are many advantages of physical stores:

1. The quality is relatively guaranteed. All foreign garbage is second-hand goods. There is a big difference between the old and the new. It's certain that you will be more relieved to see the real thing. In many cases, physical stores are more expensive than Taobao, but the author generally won't save this money, and would rather go to physical stores to buy more expensive but more attractive. Thousands of servers, if in order to save hundreds of dollars, buy a mine, have to pay off!

2. Postage problem: the server and workstation quasi system are very heavy, they are all huge things of 2mm iron sheet, weighing 20-30 Jin per minute, and the corresponding postage may be over 100, or even hundreds.

So where is the physical store? Generally speaking, coastal cities are gathering points, which are also known as big ship ports. And the mainland is more Chrysanthemum tight.

For example, Shenzhen, Guangzhou, Tianjin and Beijing all have a large amount of foreign garbage.

What about those students in the mainland? At this time, I suggest that you choose the motherboard or single channel quasi system. The postage of these things is relatively conscience.


When it comes to picking up rubbish, we can't help but talk about Baidu's stick bar graph Latin.

I also come from tulin bar.

Tulin bar, known as the largest garbage man gathering place in China, has a history of 771 to 775, exploring E3 civil value, 149 World War lol and so on.

For a long time, there were many experts in the tulin bar. It should be noted that the 771 to 775 event in that year directly caused the 775 CPU to crash, and even changed the design of pins in the later generations of Intel.

People are afraid of famous pigs and strong ones. It is because of these events that a large number of Xiaobai, fenzi and even JS are pouring into the picture bar. Now, the total number of people in the picture bar has exceeded 100W, but there is a phenomenon of starting everywhere, JS flooding, taking advantage of the cheap, turning over the car and spraying each other.

Many people have nicknamed it "Tujia bar" and "the second card bar". It has completely changed its flavor and lost the initial atmosphere of technical discussion. The first batch of "Tui bar" and "diving" are basically entertainment bars.

I belong to one of the earlier group of pushpins, a little bit of looking at the development of the bar, and finally degenerated to today. It has to be said that this matter is worth our thinking. Let's see if even freebuf has this trend?

Well, it's a partial question.

Other terms

Two way: the technology of installing two CPUs on one main board at the same time, in addition, there are four ways and so on. As shown in the figure, it is a two-way main board:

ES (QS) version CPU: refers to the engineer sample, which is the version used for testing during CPU development, and QS is the later (improved) version. These CPUs often have bugs, but the price is very low, with a considerable price performance ratio. We will focus on the discussion later. As shown in the figure is an ES CPU:

1U, 2U server: refers to the thickness of the server. In order to plug the cabinet, the shape of the blade server is uniform, with the thickness of 4.445cm as the basic unit. 1U is 4.445cm, 2U is 2 times of 1U is 8.89cm. It is recommended to give priority to 2U when playing the server, because full height graphics card can be plugged in, and only half height graphics card can be plugged in 1U. As shown in the figure, 2U blade server:

JS: it doesn't mean JavaScript. It means profiteers. I don't know why. Everyone likes to call the peddler of computer city JS, whether he is a real JS or not. I'll talk about this.

Server CPU overview

In chronological order, server CPUs can be divided into the following series:

Here we only talk about the Intel family's server CPU (for a meal, you can call me.)

Only a few of them are listed. Some like 1356 are not worth discussing.  

LGA771 series:

Early servers, tears of the times, performance is not satisfied with the current use, there is no egg to use, the only thing is that 771 hard to 775 is interesting (by hard to install the server CPU on the home motherboard), considering the stability, performance, memory rate, North Bridge fever . let's not talk about it. If you are interested, search e5450 directly.

As shown in the figure, the CPU has been modified from 771 to 775:

Lga1366 series:

The eliminated server accessories of the previous generation, corresponding to the desktop level i7, have high cost performance and are the first choice for low budget toss.

The main board is generally around 300-600 (two-way main board).

CPU is divided into two series, 5500 family and 5600 family.

The former is 45nm technology, with a maximum of 4 cores and 8 threads. Foreign garbage has been widely distributed.

The latter 32 nm process, with a maximum of 6 cores and 12 threads, will be launched one year later, so it has not been distributed yet, and the price is relatively high. It is expected that a large ship will come ashore in recent months.

There are four internal series: l, e, x, w (prefix). The first three are low, medium, and high power consumption (power consumption is proportional to frequency). Unless special requirements, it is recommended to directly go to X (high power consumption) series. W series is very high frequency, it can only be used for single channel motherboard.

W series uses less, because it can use W series single channel motherboard, often can not use RECC memory (to be discussed later).

It is suggested to give priority to 5500 series high frequency, such as x5550 and x5570, in non special cases.

I hate that 1366 series doesn't bring native sata3 and USB3.0. If you want to use SSD, please pay attention to it.

Lga2011 (E5) series:

Now one of the most valuable series, let me say three words: E5! (๑≧∀≦๑)


It belongs to in-service server accessories (V1 series is about to be eliminated), with high performance and new technical support. It is the first choice for pseudo local tyrants.

The main board is generally 800 + (one-way), but it's on the edge of the ship's landing, so it's estimated that the price will fall soon.

The CPU is divided into V1 and V2 versions, and the main board is universal.

V1 has been distributed, now the price of some models has been more conscientious, can be considered.

V2 version is in service but will be eliminated soon, which belongs to the next batch of foreign garbage.

The interior is divided into e5-1600 family, e5-2600 family and e5-4600 family.

Corresponding to single channel, double channel and four channel respectively (downward compatible).

For single core performance, the 1600 family has high main frequency (above 3.0GHz) and few cores (4c8t).

For multi-core performance, the 2600 family has a low main frequency (most of them are below 3.0GHz, but it can also be up to 3.0g through RIF Technology), and a large number of cores (most of them are 8c16t).

4600 families have few goods and are expensive, so there is no discussion value.

Lga2011v3 series:

Note that 2011 and 2011 V3 are totally different interfaces.

Most V3 mainboards in 2011 are x99 chipsets, which belong to the current popular series in the civil market. The advantage is that they can easily be sold to new mainboards.

The 2011v3 series includes E5 V3 and V4.

It belongs to the latest batch, especially V4, which was just launched last year. It's very expensive. It's only for Zhen and tuhao to choose. Now most of the ES versions are in circulation on the market.

Accessories (memory, hard disk) Introduction


Server memory can only be used on the main board of the server, and the price is extremely low, which is a big reason for the high cost performance of assembling the second-hand server.

Here we mainly discuss RECC memory of DDR3.

RECC memory, full name of register ECC memory, in which ECC refers to error checking & correcting, that is, validation memory, with higher stability than ordinary memory.

However, due to the huge amount of demolition, they are all sold by Jin, and the price is generally around 35 yuan / 4G, with a high cost performance.


Let's talk about pure ECC memory by the way.

The effective particles of pure ECC memory are less than RECC, and the cost is lower than RECC. The validity performance is not as good as RECC, but the compatibility is good. It can be used on any motherboard (including civil motherboard).

But it's very interesting that the price of pure ECC memory is higher than that of RECC memory, even ordinary memory, which has no cost performance. This proves that the price of foreign garbage has nothing to do with the value, which is basically determined by the supply and demand relationship.

So when you buy the motherboard (especially the single channel 1366, most of which only support pure ECC), you must pay attention to whether it supports RECC or pure ECC.

Another point is the heating capacity of RECC. Its heating capacity is greater than that of ordinary memory. It is recommended to insert one memory slot apart to leave room for heat dissipation. If you want to fully insert it, it is better to install a fan for heat dissipation.

Note: the industry is heavily monopolized, and the upstream says it will be as much as it wants. Recently, RECC has suddenly skyrocketed (I think it may be related to 5600 series and 2011v1 series coming ashore soon). When it reaches 75 yuan / 4G, it should be the peak in history. It can be said that it has lost the cost performance ratio completely. It is suggested to buy again after a while (maybe about a month).

The default value is 35 yuan / 4G.

Hard disk

There is not much discussion about the server hard disk. Here we discuss the SAS hard disk and its size and hard disk array.

SAS hard disk

The first is dedicated hard disk for server, namely SAS array hard disk.

This kind of hard disk has a very high revolution, generally 10K, or even 15K (generally, the hard disk is more than 7K), and the reading speed is faster than the ordinary hard disk. But it also causes high heat and noise.

If you want to buy SAS hard disk, based on the life problem, I suggest that you buy a new hard disk directly, which is more expensive than ordinary hard disk, but cheaper than SSD, such as 300g / 400 yuan.


One more thing to note:

The server hard disk is divided into two sizes: 2.5 inch and 3.5 inch. If you are playing with the server quasi system, pay attention to whether your server disk is LFF or SFF, and pay attention to the size of the hard disk before buying.

Raid hard disk array

Raid hard disk array is composed of many cheap disks, which are combined into a large capacity disk group. The bonus effect of using individual disks to provide data can improve the performance of the whole disk system. With this technology, data can be cut into many sections and stored on each hard disk separately.

Simply put, the capacity and speed of multiple hard disks are superposed. For example, two 500g and 120m / s hard disks are combined into one 1t and 240m / s hard disk to obtain the maximum speed.

Raid is divided into Raid0, Raid1, Raid0+1, Raid2 and so on.

Here we mainly discuss raid0 and RAID1.

RAID 0 is the most original hard disk array, which is pure speed and capacity, not delicate and not artificial. However, its disadvantages are poor stability and error correction ability. Once a hard disk is booted, it means that the whole raid is scrapped and all data cannot be recovered.

So when building a raid, I don't recommend using a second-hand hard disk. I don't think we should save anything else.

To ensure stability, people invented RAID1.

In fact, the principle is very simple: block the backup array and write two copies of data at the same time. Even if one array is hung, the backup array can continue to read and write, with good disk redundancy.

Although this is absolutely safe for data, the cost will also increase significantly. The disk utilization rate is 50%. For example, if two 1t hard disks are used as arrays, in fact, only 1t of space can be used.

When saving key data, you can consider RAID1. In general, you use raid0. As long as you don't need to report the wasted hard disk quickly, the probability of the raid0 boom falling is very small, so you don't need to worry about it.

In addition to 0 and 1, there are also RAID 0 + 1, 2, and so on, which are not discussed here due to space limitations. If you are interested, you can have a look at Baidu Encyclopedia. The content is surprisingly accurate (this may be the worst time that Baidu Encyclopedia has been hacked ○ RZ):


Well, so far, the author has introduced most of the basic knowledge, I believe you have a basic framework for second-hand servers.

This article is too long (now 22 pages), so it's divided into two parts.

In the next part, the author will focus on practice and discuss the common generation CPUs and their corresponding motherboards and quasi systems.

Finally, I will discuss some common accessories and configuration list.

Of course, there are also our 800 build 8-core server practice!

*Original author: push, eax, this article belongs to freebuf original award program, reprint is prohibited without permission