DIY desktop graphics card selection sharing

Recently, Aunt Zhang's DIYERs have become more active, and they have also caught up with the big events. I also come to join in a lively event. I would like to write my own graphics card purchasing principles. If there is anything wrong, I hope you can correct me. Because it was seen that the essay activities were relatively late, and the text was required to be revised once, it appeared that the essay did not catch up with the essay of the public survey activities. It was a pity that we hoped to pick up the character for the next public test.

Written in the front: This article mainly wants to discuss the choice of desktop discrete graphics in DIY, set the core display, notebook MXM card is not in the scope of discussion, and due to the explosion in recent years has eliminated the low-end graphics card, some nuclear Even HEVC can be hard-coded, so what you can buy in the market is actually a mid-range or higher-end graphics card. It's a bit like the feeling that an integrated sound card has driven an independent sound card out of the chassis. If it is only processing documents, the demand for graphics processing capability is not high. The user suggested not to buy a discrete graphics card.

Select video card to see what

If you want to get value for money, you must first work hard to understand that you know something that will not buy too much. Naturally, researchers at the University of Chicago publish articles in the Journal of Consumer Research, proving to be consumers. Time is a fatal factor in decision making . It takes a lot of time to study theoretically and it is king. If you want to buy a cost-effective graphics card, you must first understand the graphics card and know which elements determine the price/performance ratio.

Often see a lot of small white users can not buy what they want because they are not experts, such as some time ago, there is a sister paper said to see a certain laptop memory is very large to determine this video card is good; or high frequency graphics card performance is better than The frequency is low; or A card game performance is not good but the video is more exciting, N card game is smooth but the screen is not OK; there are no questions asked why my 980ti play H.265 is not as good as 950 smooth.

This all reflects the same problem. Consumers have not done enough homework and often do not know in advance what they want to buy. They do not know what to buy afterwards. Of course, we cannot ask every consumer to be very familiar with all walks of life. It is not realistic. So how do we choose the video card?

Selecting the right GPU is the key

My personal opinion is that the GPU is absolutely dominant in the selection of graphics cards, and its importance is more than 80%. The rest is in turn the brand, memory, power supply module, component performance, heat sink, etc.; that is to say, the GPU is basically stable, although the public version of the card is not particularly large, but the graphics card It is still a product with a high degree of homogeneity. The price and performance of a product with the same specifications from different manufacturers are almost indistinguishable. A 10% increase is a huge factor. How to select a GPU? Many students were confused at this time. For example, in my previous article, my classmates wanted to learn AE. Finally, the game was limited to 950 games. If the budget is adequate, then Leadtek’s m4000 card should be selected. This is buying low, unable to meet demand

Also, as mentioned above, I would like to use H.265 with 980ti to solve the problem. This is what I bought. Only 950 and 960 support hard solutions, and more advanced ones do not support it. This is what I have done. Performance has been achieved but the function is not satisfied. This time, we must conduct further analysis on the GPU.

We all know that when choosing a CPU, it depends on three main factors: architecture, instruction set, and cache .

I personally believe that the selection of GPUs should also follow similar principles: architecture, functionality, and caching .

Moreover, the proportion of the architecture in the GPU should be even more than that of the CPU. Nvidia's microarchitecture has been upgraded to Pascal, and the agricultural enterprises have also reached the Polaris. However, since the GTX7 series, the N family has clearly outranked the agricultural enterprises in the architecture. The advantage of the agricultural enterprises in the process is also not great. The improper wording of an Nvidia engineer quoted here is that “ AMD has two years of lead in hardware and 10 years in software ”. Plus I am also an N dog. I bought 6600GT 10 years ago, so I generally push N cards more.

Structure

Since the architectures of the A and N cards are completely different, although R600 seems to have published more detailed architecture information, the N card has not released too many details, but due to limited space, I will only force the recent architecture of the N card. Vertical comparison, due to limited space or from the recent several architectures, if it is really like the Xiaobian recommendation, from the beginning of the Geforce era is afraid to issue a book, so 6600GT I will not be on, but also because of the recent 3 The architectures of Fermi, Kepler, and Maxwell are in fact the same strain, and the design ideas should be the same people;

Let's take a look at Maxwell here. This architecture is used in the 750, 750ti, all 9 series, or first look at my own outdated 750ti . Last year, the buy price of Double 11 was 789.

Although the graphics card is a bit old, it is still the popular Maxwell architecture, codenamed GM107, and M is Maxwell, which is basically the lowest end of the architecture.

Theoretical test 3Dmark11 score 4198; actual test run zombies 3, radiation 4, can open full effects, COD12 not, open AA or card.

The macro can be seen that the GM107 core has five groups of SMM, that is, 640 stream processors, and the back end has 16 ROPs and two groups of memory controllers. Note: The Shaders here are a very important parameter and have a large impact on performance under the same architecture.

Each SMM in the Maxwell architecture has only 128 stream processors. Each group has its own multi-body engine 2.0 and eight texture units, and each SMM is further subdivided into four logical control cores. There are 32 stream processors in the logic control core.

First compare with the older generation.

In the age of Fermi and Kepler, each group called an SMX, which is called "Streaming Multiprocessor Extreme", and Maxwell was SMM, which is called "Maxwell Streaming Multiprocessor". Here it is obvious that Maxwell's core performance is increased by 35% and its energy consumption ratio is doubled. , is a huge improvement, which directly led to only 60W TDP 750ti, while many GPUs are three-digit.

Compared to his successors, GTX950 , just this year, 618 just helped his classmates to install one, buying price of 969.

You can see that the code has changed from GM1XX to GM206, which means that although everyone is a Maxwell architecture, it has been upgraded from the first generation of Maxwell to the second generation of Maxwell, and it can be said to be a reborn change.

The above is the GTX950 , the following is the GTX960 , there are 2 GPCs, you can see that the 950 is the castration version of the 960, less 2 SMM, chip code is GM206, so the number of stream processors is reduced by 256, the texture unit is reduced 16, but fortunately the 32 ROPs of the original GM206 are preserved, which still gives it the same pixel filling performance as the GTX 960. The 750ti has only one GPC and the ROPs are only half the size of the GM206. This is a huge gap, which is larger than the gap between the stream processor and the texture processor, which is a problem in high-resolution displays.

In other words, the GTX960 should have 8 sets of SMM units and 1024 CUDA cores; the GTX950 has two sets of SMM units culled, and the remaining 768 CUDA cores, and the corresponding texture units have been reduced from 64 to 48, and the memory is wide. It is still 128bit, the memory frequency is reduced by 100MHz, and the memory bandwidth is slightly reduced to 105.6GB/s.

Theoretical test 3DMark11 run 5825, COD12 is still very card, no way, but look HEVC is cool.

Let's take a look at the GTX970/980

I have never touched these two cards, but there are many data. It can be seen that their core codes are GM204, which are 4 GPCs. The GM204 core used by the GTX 980 has a total of 16 SMM units, and the CUDA core is 2048. GTX The 970 has 13 SMM units and 1664 CUDA cores.

Summary: This is not to discuss TATAN and other special cases, we can see that the key to the entire Maxwell architecture ahead of the previous architecture is that SMM leads the previous SMX, and the difference within the Maxwell architecture is mainly the difference between different combinations and quantities. The architecture has replaced the previous Vertex Shader and Pixel Shader with Unified Shader, and later changed it to CUDA Core. This is exactly how it works. Although the number of CUDAs contained in SSM has been reduced from 192 to 128, it has increased because of changes in the way it works. More logic control, in fact, a lot of efficiency, this is to use the updated architecture to improve efficiency, if you increase the frequency this time can improve performance, but be aware that the frequency can only be compared in the same framework, for example, you take A card and N card do not depend on the frequency to judge the performance, even if it belongs to a grade of graphics card, in addition, Maxwell belongs to a slightly favored OC architecture this year, I bought the OC version, after the performance is insufficient. You can try to see if you bought a sculpture or a mine.

Features

Someone may ask here, why look at the CPU to see the instruction set, see the GPU do not see it? I personally think that because GPU instructions can be considered more like CPU SSE instructions, CPU instructions are much more complex than GPUs, internal RISCs are used, and GPUs are more of a RISC instruction set. GPUs can be said to be heterogeneous. Environment, there are several ISA, Shader does not interrupt the abnormal service, Shader can be designed and CPU or SSE completely different, the CPU has a lot of instructions for the privilege registers, and the GPU has a lot of dedicated texture instructions and merge access instructions and On-chip memory instructions, such as memory access type acceleration instructions, are actually used for control of the CPU. It wants to connect with various peripherals, then receive various signal events, and then make corresponding processing. The GPU focuses on calculations. Its architecture is characterized by waterfall-like execution of tasks, so the use of "streaming processors" is very image. If you don't understand such a description, it doesn't matter. You can program directly on the CPU, and you need to use the Driver and API to deal with the GPU.

What features do we want to see?

Or does Maxwell give chestnuts, what are some of Maxwell's worthy features? Of course, MFAA, DSR, and H.265 encoding, but not every Maxwell is available, DSR if you use the new driver is basically all Maxwell can have, including my own use of 750ti; MFAA is only the second generation Maxwell I know, that is, the 9 series of first look, for example, I get GTX950 for classmates, playing games more cool; and H.265 is more special, look at the following figure

You can see that only 950/960 can decode H.265/HEVC, and even more advanced 970/980 can't. They can only encode. Why? Because 950/960 came out later, although the specifications were low, new features were added. If you want to see HEVC, this is still worth considering.

In addition, some students often judge the difficulty of decoding a video to determine the difficulty of decoding a video when judging the difficulty of decoding. This is only valid under the same specifications.

The left side of the above picture is H.264, and the right side is H.265. The area with small amount of information (the color change is not obvious, such as the red part of the car body and the gray part of the ground). There are fewer words, and the macroblocks where the details are more (tires) are correspondingly smaller and more numerous, and the encoded codewords are more, which is equivalent to the key encoding of the images, thereby reducing the overall The code rate increases accordingly.

In other words, many videos that do not seem to have high bit rates will probably be more strenuous if the hardware does not support H.265 decoding. This is what I said at the beginning, and some people say that 980ti cannot decode HEVC.

Brand and other factors

It should be said that the graphics card is a highly homogenized product. The difference between different brands is not too big. Even the same German system is not too bad. The process is very similar to the motherboard. Almost every manufacturer that can produce motherboards is Can produce graphics cards, do you think the motherboard manufacturers are usually good manufacturers of graphics cards, Asus, MSI, Gigabyte are big manufacturers.

If you expand the scope of the selection of the core AIC manufacturers, although now with the German system also entered, and the German system is now also acquired the past; if you choose A card is also the same reason as much as possible to find AIB's big factory.

The same GPU memory specifications are basically the same, usually a high memory generation, because the memory frequency requirements are higher, now the memory is already GDDR5, memory or DDR4; memory size is usually similar, if your display screen resolution is particularly high You can also consider choosing a larger memory, otherwise the meaning is generally not great.

If you choose the same 750ti TDP with only 70W, then you do not need to care about the radiator, the basic manufacturers have to pass, if you choose the TDP is relatively high we must look at the radiator a little to see if the heat pipe fan is not So what happened.

purchase

It is not recommended that small white buyers go to Computer City. Now that Computer City is in a dismal business, it is a case of a slaughter. If you make money, you don’t understand the money. The budget is enough to go to Jingdong. It’s not enough to find Taobao’s reliable store. Transparent, generally the price difference is not very large, Jingdong also often have activities, make enough preparations for daily shopping cart to see it on the line.

to sum up:

If you are not very knowledgeable, try to choose the major roads, big brands, what to sell, and what to choose, otherwise there will be no place for people to ask questions.

Try to choose the advanced architecture of the graphics card, often representing updated features and a longer life cycle.

The brand is mainly responsible for the after-sales service, and has little impact on performance under the same specifications. It is not recommended to pursue the brand.

A penny a minute, now the low-end graphics card has been eliminated, graphics performance has been excess, more than 2000RMB graphics card practical use for the average person is not big, unless you are doing scientific calculations on CUDA requirements, but this kind of general I suggest to buy Tesla M40 computing card, so as not to delay your work, if there is no special demand does not recommend the pursuit of high-end graphics card, if you only deal with general document types of things, it is not recommended to go on the Internet alone, after the nuclear display Will rule in the low end. Finally, I hope everyone can buy a satisfactory graphics card.

reference:

Reference 2:

Speaker Cable

Dedicated audio cable,Horn line power amplifier line,Speaker audio cable,Audio dedicated cable

Jiangsu D-Bees Smart Home Co., Ltd. , https://www.cI-hometheater.com