Shulou(Shulou.com)11/24 Report--

Just last week, Moore Thread held its fall 2022 conference and launched MTTS80, the first domestic graphics card product that supports Windows environment and DirectX graphics interface, a new multi-function GPU chip Chunxiao, MTT S3000 for server applications, and MCCX, a meta-computing all-in-one machine.

At first, the author thought that this would be a "PPT conference". Because this time the step of the Moore thread is really too big. But unexpectedly, just a week later, the MTT S80 was actually placed on the CTOnews.com desktop, and it can be used under Windows when installed with the host computer, without complicated debugging.

In this article, let's take a look at this MTT S80, which is a step forward for the development of domestic graphics cards. The test configuration is as follows:

Design Moore thread MTT S80 packaging design is very unique, the above traditional Chinese painting style line pattern highlights the selling point of its domestic graphics card. This is also the first time that CTOnews.com has tested a domestic graphics card, which is of commemorative value.

In addition to the graphics card body, there is a very simple instruction manual and a double PCIe 8Pin to CPU 8Pin cable in the package. The reason why the manual is so simple is that it is installed in the same way as a normal graphics card. Install it, open Windows, install the driver, and finish it.

The design level of the MTT S80 graphics card body is quite high. The overall design is square and full of metal style. The shell adopts an integrated design, and uses aluminum alloy die-casting + CNC technology, which greatly improves the overall structural strength of the graphics card, without the need for the graphics card bracket and do not have to worry about deformation. The heat dissipation part is designed with 3 fans, and 2 8cm fans plus the middle 7cm fans constitute a central symmetrical overall layout.

The outer edge of the fan on both sides is wrapped by two arcs, which is inspired by the hyperbolic function, which is common in mathematics, and has a sense of design when compared with the middle circular RGB fan. The three groups of fans all support intelligent speed regulation, which can not only ensure the stable operation of the GPU, but also provide a quiet experience.

The backplane is protected by a whole piece of metal, and there is a Moore thread LOGO in the middle. The right vent will be lit after it is powered on, which is very cool.

The coolest thing is the orange halo in the middle, which is lit like a thin crater, bringing endless energy.

The S80 dense heat dissipation fin can be seen from the side of the graphics card, and four 6mm heat pipes are used to run through the heat sink as a whole to help transfer heat from the GPU chip and video memory to the heat dissipation fin as soon as possible.

The best design is the 8Pin power interface on the side, which will require a larger chassis to be compatible, but it also makes the front of the chassis more concise and beautiful.

Side interface part, using the current high-end graphics card only equipped with three DP1.4a and a HDMI2.1 interface, can support up to 8K video output.

Finally, it should be noted that the MTT S80 is one of the first graphics cards to use the PCIe 5.0interface, and it is also a graphics card that supports the PCIe 5.0x16 interface, which means that it is best to match the newer motherboard to achieve the best interface performance. So Moore Thread JD.com flagship store will choose 2999 to build an Asustek B660M motherboard to sell.

Architecture parsing Moore thread MTT S80 is loaded with a multi-function GPU chip based on MUSA architecture, "Chunxiao". Compared to Moore Thread's "Suti" released in March this year, the four built-in computing engines of "Chunxiao" have been fully upgraded to support graphics and image rendering, 8K video codec, AI training and reasoning, general computing, GPU virtualization, physical simulation and other functions.

In terms of core parameters, MTT S80 is based on TSMC 7nm process and has 4096 MUSA cores, main frequency 1.8GHz, 16GB GDDR6 video memory, video memory bit width 256bit, 22 billion transistors integrated in the core, built-in MUSA architecture general computing core and tensor computing core, which can support FP32, FP16 and INT8 calculation accuracy.

We also disassembled the MTT S80, the whole card is very easy to disassemble, unscrew all the visible screws to remove the backplane and bezel. The internal workmanship is quite regular, the video memory is 8 Samsung GDDR6 flash memory, each 2GB, constitute the large video memory of 16GB.

The core code is SD102AA-500, and the GPU chip Chunxiao based on Moore thread is built.

The most special thing about MTT S80 is that it is the first GPU in China to support Windows environment and DirectX graphical interface. Moore Thread said at the press conference that at present, the Windows driver of MTT S80 has built-in MUSA DirectX Driver modules, and has completed adaptation to more than a dozen games, including Diablo 3, League of Legends and Crossing the Line of Fire, and there are more games to run, but they are still in the process of adaptation. However, as to whether it is really what it says, let's take a look at the actual measurement.

Theoretical performance first of all, let's conduct a theoretical performance test. However, before testing, we found that MTT S80 does support Windows and DirectX environments and can support DirectX 11 at the hardware level, but the driver has not yet completed the development of all functional modules, so it only supports DirectX 9 at present. At present, most of the running software is based on DirectX 11Univer 12. So we can't carry out the regular test, we can only find another way.

In Windows, there is a software that can test the performance of DX9-Unigine Valley BenchMark 1.0. in this software, MTT S80 scored 2302 points.

When we look up the ranking on Unigine's official website, we can see that MTT S80 can reach the level of GTX 1060 6G in this project.

Pixel filling rate and texture filling rate are also important indicators to evaluate the performance of graphics cards. The pixel fill ratio refers to the number of pixels that GPU can render to the screen and write to the display memory in one second. We measured the pixel fill rate of MTT S80 using Fillrate Tester to achieve a FFP-Single texture score of 188 GPixel / s. For comparison, the pixel filling rate of RTX 3060 is 85.30 GPixel / s and the pixel filling rate of 3080Ti 3080Ti is 186.5 GPixel / s.

Texture fill ratio refers to the number of texture map elements that GPU can map to pixels in one second. We can use 3DMark 06 for testing. In the end, the highest Multi-Texturing is 170GPixel / s. For comparison, the texture filling rate of RTX 3060 is 199.0 GTexel / s. The texture filling rate of RTX 3050 is 142.2 GTexel / s. There is a wide gap between different projects because the current driver has not yet optimized CPU multithreading, so the heavier the graphics load, the better the performance of MTT S80. Once the future driver optimization is completed, the performance of MTT S80 will be further improved.

Apart from the above two tests, there are not many running software on the Windows platform. So let's switch to the Linux platform and see if we can measure some data under Ubuntu. Let's try using clpeak to test its video memory bandwidth and single-precision floating-point (FP32) performance. The final measured data are as follows: the maximum bandwidth of video memory is 365 Gbps, and the maximum single precision floating point is 13.9 TFLOPS.

What level is this? Here is the theoretical performance of the desktop-side RTX 3060 12G. The video memory bandwidth and floating-point performance of the MTT S80 are slightly higher than those of the RTX 3060.

As we mentioned earlier, MTT S80 is the first domestic graphics card that supports PCIe 5.0.Therefore, we also tested its PCIe bandwidth. We use OCL Bandwidth Test to test the uplink and downlink of the interface under Ubuntu. The maximum bandwidth for uploading and downloading is 28G / s and 32G / s, which is twice the speed of most mainstream PCIe 4.0graphics cards. It can be said that the MTT S80 is a "war with the future" graphics card.

From our tests above, the pure theoretical performance of MTT S80 can reach the level of RTX 3060-RTX 3060Ti without considering environmental compatibility. However, in the Windows environment, because the driver is still trying to adapt to the DirectX and OpenGL environment, the performance of different software is very different. It can be said that the hardware level of Moore thread MTT S80 is quite online. Although driver adaptation can not keep up with the mainstream level for the time being, it also makes a good start for domestic graphics cards.

As we mentioned earlier, the MTT S80 is the first domestic graphics card to support Windows and DirectX environments, so what is its actual gaming experience? As we mentioned earlier, MTT S80 only supports DirectX 9 environment for the time being, so we can only choose some older games that have a wide audience to test. The following games we all run to 1080p low picture quality. The first is "League of Legends", which reaches 140,150 frames, which can meet the requirements of competitive display.

If you drive to 1080p high quality, the number of frames will reach an average of about 136 frames, and you can also play smoothly.

Finally, we try 2K high-definition quality, the average number of frames can be maintained at more than 120 frames, the performance is very good.

"QQ Speed" default lock 30 frames, of course, you can play.

The average frame number of "Crossing the Line of Fire" is as high as 180 frames, which makes it possible to play smoothly.

Diablo 3 is a game demonstrated by Moore Thread at the press conference, and it is true that we can play smoothly at about 90-100 frames.

"my World" has also been adapted. But the author found that the NetEase version can not be opened, the Microsoft version can be opened directly, but the average number of frames is about 40-50 frames, which is not very smooth, but it can already be played.

Finally, let's test "CS:GO". The game is very smooth to play. We can run Benchmark to achieve an average number of frames of about 213 frames.

From the adaptation of the above games, we can see that the current idea of Moore Thread is to give priority to adapt to those national-level games with a wide audience, to improve the acceptance of domestic graphics cards, and then go back to adapt to those boutique minority games. This development idea is undoubtedly correct.

Video codec for a home graphics card, not only to be able to play games, but also to have excellent video codec ability. Moore Thread said at the press conference that MTT S80 not only supports H.264 and H.265 (HEVC), but also adds the latest AV1 codec capability, and has three DP 1.4a interfaces and a HDMI 2.1interface, each of which can output 8K and 4K pictures.

The author first tried to open a 4K online video in the tubing, the perception is very smooth, did not encounter the stutter caused by poor coding and decoding. As you can see from the control panel, MTT S80 is also normally called for GPU acceleration.

So what is its video codec performance and efficiency? We need to go back to the Linux environment, use ffmpeg tools to call the hardware codec acceleration interface of vappi, and select different formats for testing. From our test results, we can normally parallel decode multi-channel video in H.264, H.265, VP9 and AV1 formats, and realize parallel coding of multiple H.264 and H.265, as well as video transcoding between multiple formats.

We prepare a 1080p video YUV data and use H.265 for multi-channel coding. In order to increase the pressure of the encoder as much as possible, we use 9-channel parallel coding. From the test results, we can see that the frame rate of each channel is 183fps, and the overall performance is better than 1080p1600fps.

In addition, we also do some tests on the performance of the decoding, when the multi-channel pressure test decodes 1080p video, the total frame rate can also exceed 1200fps. The following is the single-channel performance of 1080p video parallel 10-channel decoding in VP9 format. You can see that the frame rate is 122fps.

It can be said that the video codec performance of MTT S80 is very strong online, and the hardware capability has been laid a good foundation. For most content consumers, they can buy it back and use it directly, and there is no pressure to watch 4K HDR videos. For video creators, the coding ability of MTT S80 hardware is also very strong. But at present, there is no editing software to adapt. According to the feedback of Moore Thread internal products, they are actively driving and API adapting with domestic and foreign video editing software, hoping to gradually meet the needs of consumer video editing in the future. Moore thread can work with some domestic editing software to promote the adaptation of editing software.

AI and computing benefit from full-function MUSA architecture, and MTT S80 can also be used in AI training. For example, developers can simply and quickly migrate existing AI models to MTT S80 through MUSA software stack; in terms of compatibility, MTT S80 is compatible with a variety of mainstream deep learning frameworks such as PyTorch and TensorFlow, and optimizes dozens of AI models such as Transformer, CNN, RNN and so on.

In our previous test, MTT S80 has strong single-precision floating-point performance, so it can show strong performance in AI high-precision reasoning with single-precision floating-point performance, which can meet the needs of scenarios with high data calculation accuracy, such as medical, financial and other application fields. For example, MTT S80 is particularly adapted to the medical domain AI open source framework MONAI, to achieve high-precision reasoning of a variety of tasks.

The largest cool techs is still CUDA on MUSA. In order to reduce the migration cost of users, Moore thread has developed a set of CUDA ON MUSA compatibility scheme for users using CUDA language. Based on the porting tool provided by Moore thread, the CUDA source code can be run on the Moore thread MUSA architecture GPU by compiling and running.

Summary: a big step for domestic graphics cards can be seen from the author's evaluation today that the hardware performance of MTT S80 has reached the level of mainstream dessert level, which is undoubtedly a big step for the whole domestic graphics card industry. However, the biggest difficulty lies in how to develop the driver later. Because of the professional nature of computer graphics, few people around the world know how to develop Windows drivers. Most of them are concentrated in western countries, and there are only a handful of professionals in China. Domestic GPU companies in the start-up stage need to quickly launch market-oriented GPU products, but the problem they face lies in the lack of talent and team honing in key areas such as chip design, underlying driver development and so on. So developing a generic GPU is by no means easy.

Even after more than a decade of nuclear display, intel, which has the largest market share, has encountered setbacks in driving development when entering the independent graphics card market, let alone for a new player who has started for 2 years. It is undoubtedly a long and difficult process for domestic GPU to be compatible with the old software ecology. We have to admit that independent innovation is a very difficult road, but it is also a road that we have to take. With a recent ban in the United States, Nvidia has to cut off the supply of specified models of GPU chips to China, and it is even more difficult for us to imagine what kind of friction will occur in the future, so we must make adequate preparations.

But today we are also fortunate to see that Moore Thread has taken the first step towards compatibility with mainstream platforms. As for the MTTS80 we have, for most mild consumers, it can be bought and plugged directly into Windows computers, and it is no problem to watch videos and play LOL. But we should also treat it rationally, we can't expect Moore thread to rise to the sky in one step and directly make mainstream-level products, so the author gives the greatest encouragement and tolerance when evaluating Moore thread MTT S80. Of course, I still hope that Moore thread can promote the adaptation of all kinds of games and applications as soon as possible, fully release this powerful core, and respond to the expectations of the entire domestic industry.

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.