Shulou(Shulou.com)05/31 Report--

Today, the editor will show you how to analyze server memory and video memory. The knowledge points in this article are introduced in great detail. Friends who feel helpful can browse the content of the article with the editor, hoping to help more friends who want to solve this problem to find the answer to the problem. Let's follow the editor to learn more about "how to analyze server memory and explicit memory knowledge".

The early memory is connected to the north bridge through the memory bus, and the north bridge communicates with the CPU through the front-end bus. From Intel Nehalem, the North Bridge is integrated into the CPU, and the memory is directly connected to the CPU through the memory bus.

Therefore, after the AMD adopts Socket FM1,Intel and LGA 1156 slot, the processors are integrated with the north bridge, the independent north bridge has disappeared, and only the south bridge is left on the motherboard.

The main contradiction of the computer system is that the CPU is too fast and the disk is too slow. So the two can not communicate directly, need to add a transition layer, this is the role of memory. Harvard architecture is a memory structure that separates program instruction storage from data storage.

Memory (Memory), also known as internal memory, is used to temporarily store operational data in CPU and data exchanged with external memory such as hard disk. All programs in the computer run in memory, so the performance of memory has a great impact on the computer.

At the end of 1996, SDRAM began to appear in the system. Unlike earlier technologies, SDRAM was designed to synchronize with the timing of CPU.

SDRAM can also be called SDR SDRAM (Single Data Rate SDRAM). Single Data Rate is a single data transfer rate. The core, Icano and equivalent time pulse of SDR SDRAM are all the same. SDR SDRAM can only read and write once in a cycle. If you need to write and read at the same time, you must wait until the previous instruction has been executed before it can be accessed.

DDR SDRAM is a dual-channel synchronous dynamic random access memory, which is a new generation of SDRAM technology. Unlike SDR (Single Data Rate), which can only read and write once in a single cycle, the double data transfer rate of DDR means that it can be read or written twice in a single cycle. When the core frequency is constant, the transmission efficiency is twice as high as that of SDR SDRAM.

Summary: DDR uses clock pulse rising edge and falling edge to transmit data once respectively, and one clock signal can transmit twice as much data as SDRAM, so it is also called double rate SDRAM. Its multiplier is 2.

DDR2 SDRAM is a dual-channel twice synchronous dynamic random access memory. The DDR2 memory Prefetch has been raised to 4bit again (twice as much as DDR), and the Imax O pulse of DDR2 is twice that of DDR.

Summary: DDR2 still uses the technology of transmitting data once when the clock pulse rises and falls (not twice), but one pre-read 4bit data is twice as much as DDR's first-time pre-read 2bit, so its multiplication coefficient is 2X2=4.

DDR3SDRAM is a dual-channel cubic synchronous dynamic random access memory. DDR3 memory Prefetch is upgraded to 8bit, that is, 8 bits of data is accessed at a time. The DDR3 transmission rate is between 800 and 1600 MT/s.

In addition, the specification of DDR3 requires that the voltage be controlled at 1.5V, which is more power-saving than the 1.8V of DDR2. DDR3 also adds two functions, such as ASR (Automatic Self-Refresh) and SRT (Self-Refresh Temperature), so that when the memory is dormant, it can also control the charging frequency of memory particles with the change of temperature to ensure the integrity of system data.

Summary: as an upgraded version of DDR2, the most important change of DDR3 is to pre-read 8bit, which is 2 times as much as DDR2 and 4 times as much as DDR, so its multiplier is 2 / 2 / 2 / 8.

DDR4 SDRAM provides a lower supply voltage of 1.2V and higher bandwidth than DDR3/DDR2. Four new Bank Group groups have been added to DDR4. Each Bank Group has the action characteristics of independently starting operation, such as reading and writing. The Bank Group group can use the concept of multiplex to imagine, and it can also be interpreted as that DDR4 can handle at most four groups of data in the same time cycle, which is obviously better than DDR3.

In addition, DDR4 adds DBI (Data BusInversion), CRC (Cyclic Redundancy Check), CA parity and other functions to make DDR4 memory faster and more power-saving, while enhancing signal integrity and storage reliability.

In 2017, Intel launched Purley, a server platform corresponding to the sixth-generation Core Skylake, using 14nm technology, up to 28 core 56 threads, 6-channel DDR4 memory, fiber interconnection channel, using UPI bus instead of QPI bus, and so on. UPI is the abbreviation of UltraPath Interconnect (HyperChannel Interconnection). The data transfer rate can reach 9.6GT/s and 10.4GT/s. It has more bandwidth and more flexibility. Each message can send multiple requests.

The three major evolution directions of memory in the future are capacity, voltage and frequency.

The capacity is getting larger and larger (4GB-> 8GB-> 16GB-> 32GB-> 64GB-> … 512GB)

The voltage is getting lower and lower (1.5v-> 1.35v-> 1.2v-> …)

The frequency is getting higher and higher (1333-> 1600-> 1866-> 2133-> 2400->.. 3200)

The mainstream memory manufacturers are divided into memory granule manufacturers and module manufacturers, and the three major memory granule (DRAM) original manufacturers are Samsung, SK Hynix and Micron. Module manufacturers Ramaxel and Kingston make memory strips (DIMM) by buying particles from particle manufacturers.

There are three different frequency indicators in memory, which are core frequency, clock frequency and effective data transmission frequency.

The core frequency is the working frequency of the memory Cell array (Memory Cell Array), and it is the real running frequency of the memory.

Clock rate is the transmission frequency of I/OBuffer (input / output buffer)

The effective data transmission frequency refers to the frequency of data transmission.

System maximum memory bandwidth = memory nominal frequency * number of memory bus bits * number of channels * number of CPU

Actual memory bandwidth = nominal memory frequency * number of memory bus bits * number of channels actually used

Actual memory band = memory core frequency * number of memory bus bits * number of channels actually used * multiplier.

Since the era of SDRAM-DDR, the data bus bit width clock has not changed, all are 64bit, but if the dual-channel technology is used, the bit width of 64*2=128bit can be obtained.

The bandwidth of a memory strip named DDR31066 at the default frequency is calculated below. 1066 refers to the effective data transmission frequency, divided by 8 is the core frequency. A memory only uses single-channel mode, and the bit width is 64bit. So, the actual memory bandwidth = (1066 + 8) * 64*1*8=68224Mbit.

Thus, if the memory works at the nominal frequency, we can directly use the nominal frequency * bit width * the number of channels actually used, and simplify the formula = 1066*64*1=68224Mbit.

If memory bandwidth is the key to exchanging data between processors and memory, video memory bandwidth is also important for video cards. GPU core is responsible for computing, video memory is responsible for data storage, the two need to exchange data frequently, which depends on the video memory bandwidth, higher bandwidth can make the video card more handy when dealing with high resolution and high picture quality.

Video memory bandwidth is determined by video memory frequency and bit width, but the actual bandwidth depends on the specific situation. At present, most of the bit widths of mainstream graphics cards are 128bit, 256bit, 384bit and 512bit, which can better determine the bandwidth or the type of video memory, which determine the limit of video memory bandwidth.

At present, the most mainstream video memory is GDDR5, of course, there was a short-lived GDDR4 before, and now there are remnants of gDDR3 video memory on the low-end market. AMD uses HBM video memory on its graphics card, which is more powerful than GDDR5 video memory and greatly increases its bandwidth.

In terms of these two, the bandwidth of GDDR5 is 32bit, the GDDR5 memory frequency of the current NVIDIA card can reach 1750MHz, it is a 4-fold rate mechanism, the data frequency is 7Gbps, and the bandwidth of a single chip is 28GB/s. At present, the frequency of HBM video memory is only 500MHz and the data frequency is 1Gbps, but its bandwidth is very high, which makes up for the lack of frequency.

GDDR5 and HBM video memory are the most mainstream video memory technology at present. At present, gDDR3 video memory is basically used by some low-end video cards of NVIDIA and AMD. GDDR5 is definitely the mainstream at present, the capacity of a single chip is gradually increased from the previous 2Gb to 4Gb, Meguiar also produced 8Gb (1GB) particles not long ago, and high-end graphics cards only need 4-8 chips to achieve 4-8GB capacity video memory, which will further promote the emergence of high-capacity graphics cards.

HBM is a rising star. At present, it is only used by AMD. The first generation HBM technology shows memory core capacity 2Gb (1 stack is 4 cores), data frequency 1Gbps, bit width 1024bit.

Video memory bandwidth = equivalent data frequency of video memory (Gbps) * total bit width of video memory (bit) / 8 = actual frequency of video memory (MHz) * data magnification of video memory (1,2,4) * equivalent bit width of video memory (64-512bit) / 8

As graphics card manufacturers are more accustomed to marking product specifications with larger and better-looking data frequencies, the above companies can actually be simpler and directly become:

Video memory bandwidth (GB/s) = data frequency (Gbps) * equivalent bit width (bit) / 8

Take the GeForce GT 720 video card of NVIDIA as an example. The bit width of the card is only 64bit, and both gDDR3 and GDDR5 video memory are supported. The typical frequency of the former is 900MHz, and the typical frequency of the latter is 1250MHz. Under the two configurations, the bandwidth is:

The bandwidth of the gDDR3:GT 720 video card is: 900MHz * 2x64 bit/8= 14.4GB/s, or 1.8Gbps * 64bit/ 8 = 14.4GB/s.

The bandwidth of the GDDR5:GT 720 video card is 1250MHz * 4x64 bit/8=40GB/s, or 5Gbps * 64bit/8=40GB/s.

Thank you for your reading, the above is the whole content of "how to analyze server memory and memory knowledge", learn friends to hurry up to operate it. I believe that the editor will certainly bring you better quality articles. Thank you for your support to the website!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.