Review of Moore Threads MTT S80 and S70 video accelerators, completely Chinese development

Last year, the Chinese company Moore Threads introduced the first gaming video cards S80 and S70, created using its own technologies. Founded in 2020, the organization has an ambitious goal to develop and manufacture a GPU using exclusively Chinese technology. During the year, the company attracted three rounds of investment, receiving significant financial injections from Sequoia Capital China, ByteDance, Tencent and other large companies. The latest investments were aimed at launching mass production of video cards and expanding the Moore Threads ecosystem. The company has reportedly established partnerships with hundreds of Chinese firms specializing in graphics and computing.

In China, many startups have emerged in recent years to develop their own GPUs, but Moore Threads has the experience and qualified staff to create its own GPU. Some of the company's employees previously worked at Nvidia, Microsoft, Intel, ARM and other large technology firms. The company has also received support from Chinese authorities by adapting its GPUs to work with local CPUs and operating systems. Currently, Moore Threads graphics cards are compatible with the major platforms preferred by Chinese users, including Intel, AMD, Loongson, Zhaoxin, as well as Windows operating systems, Kirin, Tongxin, Ubuntu and others.

Moore Threads introduced its first video cards, the S80 and S70, last fall. This is one of China's steps towards creating its own high-performance computing chips. Recent US bans on the supply of some Western GPUs to China have underscored the importance of having homegrown technologies in strategic industries, including computing and artificial intelligence.

Founded in 2020, Moore Threads has received significant investment and talent from major technology companies. They are working on their own GPUs, supported by the Chinese authorities to work together with local processors and operating systems.

Despite a number of technical and production difficulties, MTT S80 and S70 can apparently be in demand in China for watching videos and playing games that are less demanding on performance. However, these video cards are still not powerful enough to achieve world-class performance.

There is not enough information about MTT products, which makes it difficult to create a comprehensive review. Despite the lack of detailed data on MTT video cards, their appearance is of interest as one of the rare cases of new GPU manufacturers emerging. Building a competitive GPU is becoming increasingly difficult these days, and for now, it's preferable to stick with AMD, Nvidia, and Intel graphics cards.

The names of the Moore Threads S80 and S70 video cards were chosen more than a year ago in accordance with the system adopted by the company. Previously, the company had the S60 model, and later more powerful variants were released. For now, it is expected that the titles will continue sequentially.

It is difficult to talk about prices and competitors on the global market for video cards from Moore Threads due to the lack of official information and the rapid dynamics of retail prices. In China, the S80 was initially priced at 3,000 yuan ($423), but has now dropped to 1,200 yuan ($164). However, this is also considered a bit expensive. This may be interesting for the Chinese market due to potential sanctions, but Moore Threads does not promote its video cards in other countries.

Regarding the amount of video memory, for the MTT S80 with a video memory bus width of 256 bits, the company chose 16 GB, which is suitable for use in serious computing tasks. However, in games, the speed of Moore Threads video cards is so low that even when using low graphics settings, video memory capacity will not play a big role. In the case of the S70, due to the change in bus width to 224 bits, 7 GB of video memory was chosen, which may cause some compatibility difficulties with some games.

Since Moore Threads independently produces MTT video cards, they have a reference design and a three-fan cooling system that occupies approximately 2.5-3 slots in the case. They are equipped with three DisplayPort 1.4 connectors and one HDMI 2.1 connector for connecting monitors.

However, MTT graphics cards are not energy efficient and consume large amounts of power. Maximum claimed consumption is up to 255W for the S80 and 220W for the S70. This requires the use of a large cooler with three fans and one 8-pin auxiliary power connector.

The MTT S80 and S70 graphics cards are gaming variants built on different versions of the GPU codenamed Chunxiao, which is based on the Moore Threads Unified System Architecture (MUSA). GPU production began in November 2022 using the 7nm process technology at TSMC fabs. Initially, there was an assumption about the possibility of production exclusively in China at SMIC factories, but sanctions restrictions on the production of microchips using modern technologies in Taiwanese factories for Chinese manufacturers confirmed that TSMC became the manufacturing factory.

In October of that year, the US tightened sanctions against Chinese companies, banning graphics chip designers from using contract manufacturers such as TSMC. These restrictions made it impossible for Moore Threads to order the production of GPUs they developed at TSMC factories. A similar situation happened to Huawei several years ago, and now it is collaborating with the also sanctioned Chinese company SMIC. As a result of the sanctions, Moore Threads has announced some staff cuts and will likely also have to rely on SMIC's capabilities in the future.

Built on TSMC's 7nm process, this Chunxiao GPU has some truly impressive specs. The transistor count is 22 billion, which is significantly higher than the 17.4 billion transistors in the GA104 chip used in Nvidia's GeForce RTX 3060 Ti, RTX 3070, and even Nvidia's RTX 3070 Ti. The die area is also impressive at 416 mm², versus 392 mm² for the GA104. These numbers indicate the Chunxiao's high transistor density of 52.9 million/mm², which beats the GA104's 44.4 million/mm². This difference is explained by the use of different process technologies: 7 nm TSMC for the Chinese GPU and 8 nm Samsung for the Nvidia GPUs.

Chunxiao is known to include 4,096 MUSA stream processing units running at 1.8 GHz, delivering up to 14.4 teraflops of single precision floating point (FP32) computing performance. In addition, the presence of 128 tensor cores is claimed, providing matrix calculation performance in INT8 format up to 57.6 teraops.

Unfortunately, detailed data on the number of TMU texturing units and ROP rasterization units in Chunxiao is unknown, but based on the data on peak frequencies of 460 Gpix/s and 460 Gtex/s, we can assume that the number of these blocks is about 256 pieces. These numbers represent a large number of blocks compared to other GPUs.

However, the lack of detailed information about the internals of this GPU may raise questions about its efficiency and competitiveness, even given its theoretical performance.

An interesting observation regarding computing efficiency when comparing the Moore Threads MTT S80 to Nvidia GPUs. It's clear that the MTT S80, despite its impressive transistor count and matrix compute performance, is inferior in terms of single precision floating point performance to the FP32 when compared to Nvidia chips such as the GA104 found in the RTX 3060 Ti and RTX 3070 Ti. Indeed, this is an interesting observation, which indicates the different efficiency of chips depending on their architecture and configuration.

As for the Moore Threads MTT S80 architecture, it features 16GB of high-performance 14GHz GDDR6 memory coupled to the GPU via a 256-bit bus, delivering 448GB/s memory bandwidth, similar to the RTX 3060 Ti. However, an important difference between the older MTT S80 model is the support of the PCIe 5.0 interface with full x16 width — the industry's first GPU with such a high-speed interface, which, in theory, provides significantly higher data transfer capacity of up to 128 GB / s in both directions. However, the availability of tests to confirm these values in the Windows operating system was not found, and researchers have to rely on statements from Chinese manufacturers.

These specifications make the MTT S80 an interesting contender in the graphics card market, however, more extensive and robust testing will be required to confirm actual efficiency and performance in real-world conditions.

Later, the MTT S70 model was introduced and released onto the market, based on the same GPU as the S80, but with some reduced characteristics. The MTT S70 graphics card is equipped with 3584 stream processors of the MUSA architecture and a clock speed of up to 1.6 GHz, compared to 4096 processors and 1.8 GHz for the S80. This resulted in a reduction in the S70's computing performance by 3.2 teraflops. Probably, the younger version of the GPU also has reduced texturing and filling units. In addition, the S70 has a 224-bit memory interface, allowing for 7 or 14 GB of video memory, the second option selected, which seems unusual.

It is interesting to note that the MTT S70, unlike the top-end S80 model, only supports the PCIe 4.0 interface, while the S80 boasts support for PCIe 5.0. However, PCIe 4.0 in gaming scenarios may be preferable in terms of compatibility and speed for slower GPUs. The power consumption of the younger version of the video card has not decreased much: from 255 W to 220 W.

Both MTT models feature a second-generation multimedia engine that supports common video formats, including AV1, H.264, H.265 and VP9, as well as hardware-based HDR10 video playback. The S80 and S70 have four blocks that process multimedia data, useful in computer vision tasks and accelerating the encoding/decoding of video streams in major formats. Supports hardware acceleration of video stream decoding in H.264, H.265 (HEVC), VP9 and AV1 formats, including 10-bit profiles with a resolution of up to 8K or up to 32 Full HD channels at 30 FPS. Test video viewing showed GPU load at approximately 50%-60%.

Thanks to support for relatively modern display interfaces DisplayPort 1.4a and HDMI 2.1, the Chinese video cards in question support simultaneous output of information to four displays with a resolution of up to 8K (7680×4320) at 30 Hz or 1920×1080 at 360 Hz, support is also promised HDR. Well, support for multi-stream MST transport makes it possible to connect several displays to one port by combining several video signals into one stream — up to four.

Unfortunately, finding specialists who know how to create GPUs is not a trivial task, as we will discuss later. However, it is comparatively difficult to find qualified programmers capable of developing optimized graphics drivers for Windows and DirectX. It is a highly specialized field in which the number of experts is limited and most of them are located in Western countries. Even Intel, which has been in the GPU business for decades, has faced many challenges and failures in creating graphics hardware and software, not to mention a new entrant that has only been in the market for three years. Therefore, it is not surprising that Chinese companies are having… let's say, some difficulties, and this is even a softened expression.

The Chunxiao hardware GPU and MTT graphics cards look very interesting, but from a software perspective they still have a lot of work to do. The PES Graphics Card Control Center includes functions for updating drivers, monitoring the status of the video card, some settings, etc., and overall it works quite well. Support for DirectX, OpenGL, Vulkan, OpenCL and CUDA (via broadcast) is announced, but at the moment this is limited to DirectX 11 with feature level 11_1 and OpenGL 3.3. Thus, the GPU only supports older games and applications, and even there there are performance and quality issues.

Of course, Chinese developers are constantly improving drivers, and new versions often bring noticeable performance gains. However, this is due to the low starting level: the initial performance was very modest. Moore Threads continues to optimize the software for better performance and compatibility, but there is no guarantee that significant changes will occur. Currently, support for Direct3D features is limited to these features.

Diagnostic tools show the total available video memory at 32 GB (the sum of 16 GB of local video memory and another 16 GB of shared memory in RAM). However, the levels of Direct3D functionality supported by MTT video cards are limited to version 11_1, although modern GPUs support versions 12_1 and 12_0. The maximum size of supported textures is 8192x8192 pixels, while for Nvidia video cards this parameter can reach 16384x16384.

Features not supported in Direct3D 11 on MTT GPUs include AGP Texturing, Double-Precision Floating-Point, Driver Concurrent Creates, Driver Command Lists, Edge Anti-Aliasing, and Tiled Resources. While Edge Anti-Aliasing and Double-Precision Floating-Point are rarely used with this level of GPU, command lists and tiled resources could be useful, especially given their importance for Direct3D12, which is not currently supported by MTT GPUs. These functions allow you to create objects (textures, buffers and shaders) in parallel in different threads, as well as create a command buffer in a separate thread, which speeds up loading and processing of data. These features are widely used in modern Direct3D12 applications. Tessellation is also not fully supported, at least for now.

It is also important to note the lack of support for the hardware acceleration feature of the GPU scheduler (Hardware Scheduling), which was introduced in the Windows 10 update in May 2020. This feature allows for more efficient distribution of resources among different applications. With this feature enabled, the GPU can handle most of the planning, which usually works more efficiently. However, the operating system still retains control over application priorities when using resources.

Developing your own graphic architecture from scratch is a complex and time-consuming task that requires deep knowledge and experience. In the case of Moore Threads, although there is no direct evidence of how they achieved this, it is speculated that they may have licensed the architecture from Imagination Technologies.

Imagination Technologies has extensive experience in GPU development, with a portfolio including PowerVR Kyro in desktop PCs and graphics cores for mobile devices. They are famous for their tile-based deferred rendering (TBDR) architecture, which provides a good balance of performance and energy efficiency.

Imagination recently released IMG DXD, high-performance DirectX-enabled graphics engines available for licensing to third parties. These cores offer significant performance gains over previous solutions, leveraging various technologies such as increased FP16 compute performance, support for texture compression, and an integrated RISC-V core for GPU management.

This licensing model allows companies, including Moore Threads, to use off-the-shelf solutions and technologies, saving the time and resources of developing their own architecture from scratch. This is a common practice in the industry, especially for new players who are looking to quickly enter the market and produce competitive products.

With a dual-core configuration, the IMG DXD GPU provides up to 5 teraflops of processing power in FP32 format, and its capabilities in FP16 and DOT8 are several times higher. Texturing speed reaches 144 gigatexels/s, which is quite enough for unassuming gamers. Real-world applications confirm that IMG DXD performs 40% to 60% better than equivalent IMG BXT configurations. The ratio of processing power, texturing speed and filling in IMG DXD is almost the same as stated in the Moore Threads materials, which suggests the possible use of previous Imagination solutions in development.

The growing demand for high-performance and energy-efficient video cards in the Chinese market, caused by increasing Western sanctions and restrictions, makes Imagination's offer very popular. Although IMG DXD's DirectX hardware support level is currently limited to version 11_0, which is slightly lower than MTT, it can run many popular games and applications. Imagination Technologies plans to further develop its solutions and improve their capabilities, including support for new API versions such as Vulkan 1.3, OpenGL 4.6, OpenGL ES 3.2, and OpenCL 3.0.

Rumors talk about the possible use of PowerVR architecture in MTT solutions, which could serve as an example for other Chinese companies in contracts with Imagination. However, although Chinese companies have to rely on ready-made foreign blocks when creating a GPU structure, the use of their own technologies and developments is becoming an important aspect of their development, given the risks associated with increasing sanctions pressure from Western countries.

Despite the possible influence of Western patents and technologies in the GPU field, many Chinese startups are engaged in research and development in this area, mainly using Imagination or NVDLA cores. Developments like these open up new prospects for China's GPU and compute processor industries, which could lead to the development of more independent technologies in the future.

Let's discuss the practical results of using MTT video cards from different researchers. According to the Fillrate Tester, the MTT S80's fill rate reaches 188 Gpix/s, which is significantly faster than the RTX 3060 and close to the RTX 3080 Ti. This indicates a high theoretical peak fill rate, but actual performance varies.

Tests in 3DMark 06 show texturing speeds of about 170 Gtex/s, which is closer to competitors, for example, the RTX 3060, which provides about 200 Gtex/s.

However, there were few available tests for Chinese video cards under Windows. It was not possible to conduct tests using OpenCL, although the speed of FP32 floating point calculations, according to other researchers, was about 14 teraflops, close to theoretical indicators. Estimating PCIe throughput also remained impossible under Windows for MTT cards, but Chinese researchers achieved upload and download values of up to 28 Gbps and uploads of up to 32 Gbps on Ubuntu via OpenCL, surpassing most PCIe 4.0 cards.

MTT's hardware capabilities for decoding VP9 videos up to 1080p resolution showed frame rates of over 1,200 fps for 10 channels, and for H.265 encoding, it achieved frame rates of over 180 fps for ten 1080p streams. However, the adaptation of software to these solutions remains insufficient.

In 3D performance, the MTT S80 should theoretically reach the level of the GeForce RTX 3060, but due to unoptimized drivers, performance even in games supported by Chinese GPUs is lower than expected. Some test results do not correspond to the stated theoretical characteristics of the MTT S80/S70, which is likely due to imperfect drivers.

MTTs exhibit better performance in applications using DirectX 9, hinting at a legacy architecture for which no optimization has been done. DirectX support for Chinese video cards is gradually expanding, but does not yet cover Direct3D 12, which is used in most new games.

There are significant problems with MTT graphics cards in gaming environments. The software side is under constant development, and the drivers do not support many standard graphics APIs such as DirectX, which causes performance and quality issues. Technical bugs in the GPU hardware require software workarounds, which negatively impacts performance. Although MTT's architectural capabilities promise a lot, in reality their gaming performance is comparable to that of a GTX 1050 Ti or GTX 1650, which is clearly not enough for modern games.

Despite the promises and theoretical indicators, MTT video cards are not yet able to compete in the gaming solutions market. Their use in games is limited due to driver and API compatibility issues. Although the hardware capabilities allow us to expect more, in practice the current performance remains insufficient even for the Chinese market, especially considering the initial prices of these video cards.

Hope for improving the situation comes from the experience of Intel, which also faced similar difficulties when entering the graphics card market. Perhaps further development and refinement of drivers will lead to improved performance and compatibility of MTT video cards. However, their use in games currently seems unjustified, and actual performance falls far short of expected levels.

MTT graphics cards are likely better suited for computing tasks, including use in neural networks and deep learning, rather than gaming scenarios. The MUSA architecture and the company's full software stack (MUSA) provide tools for developers to port programs to MTT graphics cards compatible with deep learning frameworks such as PyTorch, TensorFlow, PaddlePaddle, OneFlow and others.

Based on the theoretical capabilities and functional modules included in the MUSA software stack, MTT video cards should have good performance in such computing tasks. Their maximum computing performance in single-precision floating point operations and a special 8-bit INT8 format is high enough for working with neural networks.

There is information that MTT solutions are well adapted to work with the MONAI artificial intelligence platform in the field of medical research. There is also evidence of high performance in OpenCL under Ubuntu. However, since the data is based on public tests, without direct use and calculations, it is difficult to independently verify these claims.

CUDA on MUSA, developed by Moore Threads, aims to make it easier for existing CUDA users to migrate from Nvidia graphics cards to MTT graphics cards by porting the CUDA code to code compatible with the MTT architecture, most likely using OpenCL or similar tools.

Porting CUDA source code to MTT GPUs is said to be relatively simple, requiring the code to be ported and recompiled to run on the MTT hardware. However, without conducting actual tests or using this technology directly, it is difficult to say how effective and simple it is in practice.

It is important to note that while CUDA on MUSA may provide a convenient way to port code from CUDA to the MTT architecture, the efficiency and performance of this port may vary depending on the complexity of the code, the CUDA features used, and the optimizations that may be required to adapt to the new architecture.

Moore Threads Technology (MTT brand) was founded in 2020 in the People's Republic of China. Headquarters in Beijing. Initially it was created as a startup company for developers of new graphics processors and video cards based on them. The formal leader is Zhang Jianzhong, former vice president and head of Nvidia's office in China. Over the three years of its existence, the startup received billions of dollars of government support from the Chinese government, as a result of which a year later the company announced the first successful development of its own GPU, on the basis of which the company’s first products for computing needs were released under the unified ideology of creating “metacomputers” for computational processes of digitizing everything physical peace and physicalization of the digital world. A unified system architecture, MUSA (MT Unified System Architecture), was created. The number of personnel is unknown.

Research objects: 3D graphics accelerator (video card) MTT S80 16 GB 256-bit GDDR6 and MTT S70 accelerator 7 GB 224-bit GDDR6.

Externally, the cards are exactly the same.

It is clearly visible that both cards have exactly the same PCB. The only difference is the capacity of the installed memory chips and the absence of one chip in the S70, as a result of which the memory volume has decreased to 7 GB, and the width of the memory exchange bus has decreased to 224 bits.

MTT S80 uses SD102AA-500 chip, release date unknown.

MTT S70 uses SD102AA-400 chip, release date unknown.

The chips have different production markings, so we can assume that the crystals were produced at different factories.

The total number of power phases for both cards is 8 (6+2).

The core power supply circuit is marked in green, and the memory in red.

The 6 phases of the core power supply are controlled by two DRV8305 PWM controllers (Texas Instruments), each of which is designed for a maximum of 3 phases. They are located on the front of the card.

The 2 power phases of the memory chips are managed by exactly the same third PWM controller.

The power converter for the core and memory chips uses DrMOS transistor assemblies — in this case AOZ5311 (Alpha&Omega Semi), rated for a maximum of 55 A.

There is also a Texas Instruments controller for monitoring (monitoring voltages and temperatures).

The cards are equipped with backlighting, but it is monochromatic and uncontrollable, so there is no backlight controller.

Both cards are equipped with a Realtek RTD2175 controller, which converts the DisplayPort signal to HDMI 2.1 (for the only such video output).

The power consumption of the cards in tests reached up to 170 W for the MTT S80 and up to 152 W for the MTT S70 (which roughly corresponds to their stated maximum consumption, taking into account the fact that during the tests we saw consistently low GPU load, far from 100%).

Power is supplied to both cards through the side end. At the same time, the MTT S70 has a regular 8-pin PCIe 2.0 connector, and the MTT S80 also has an 8-pin, but EPS12V (these are installed on motherboards to power the CPU).

The MTT S80 card comes with a power adapter to EPS12V from two 8-pin PCIe connectors (usually power supplies have only 2 EPS12V tails, and often both of them are used).

In order not to confuse the required connector, the cards are equipped with corresponding stickers on the power connectors. However, we received the MTT S80 as part of an already assembled PC, with the power tail and adapter already inserted, so there was no sticker on it.

Management of the cards is provided using the proprietary MTT PES (Perfect Experience System) utility. PES allows you to see monitoring in all its glory and control the output of graphics on multiple monitors. There is no way to control GPU frequencies and operating modes. It is very important to enable automatic tracking of driver updates in the settings, because only half of the versions are posted on the website in the form of separate packages, the rest are downloaded via PES in the form of updates.

The MTT family already understands the beta version of FurMark 2.1, but GPU-Z and HWinfo still cannot work with such cards.

The basis of the CO is a relatively massive two-section nickel plated radiator with heat pipes that distribute heat along the fins of the radiator.

The tubes are soldered to a huge copper plateau. Memory chips and VRM power converters are cooled using the same huge heat sink (through thermal pads).

The back plate serves as a PCB protection element, enhances the rigidity of the structure, and also helps in cooling the power elements of the GPU power circuit (it is pressed to the back side of the PCB through a thermal pad).

A casing with three fans is installed on top of the radiator (side fans — ∅100 mm, central — ∅85 mm).

The fans always rotate, regardless of the load.

Temperature monitoring

MTT S80

After a 2-hour run under load, the maximum core temperature did not exceed 52 degrees, which is an excellent result. The power consumption of the card reached 170 W.

The maximum heating is near the GPU and at the PCIe connector.

MTT S70

After a 2-hour run under load, the maximum core temperature did not exceed 46 degrees, which is also an excellent result. The power consumption of the card reached 152 W.

The noise measurement technique assumes that the room is soundproofed and muffled, and reverberations are reduced. The system unit in which the noise of video cards is studied does not have fans and is not a source of mechanical noise. The background level of 18 dBA is the noise level in the room and the noise level of the sound level meter itself. Measurements are taken from a distance of 50 cm from the video card at the level of the cooling system.

Measurement modes:

• Idle mode in 2D: Internet browser loaded with iXBT.com, Microsoft Word window, a number of Internet communicators
• 2D mode with movie viewing: SmoothVideo Project (SVP) is used — hardware decoding with insertion of intermediate frames
• 3D mode with maximum load on the accelerator: FurMark test is used

The noise level rating is as follows:

• less than 20 dBA: relatively silent
• 20 to 25 dBA: very quiet
• 25 to 30 dBA: quiet
• 30 to 35 dBA: clearly audible
• 35 to 40 dBA: loud but tolerable
• above 40 dBA: very loud

MTT S80

In the no-load state in 2D mode, the temperature remained at no higher than 24 °C, while the fans ran at 1100 rpm, which provided a noise level of 22 dBA.

When watching a movie using hardware decoding, the temperature and fan settings remained unchanged.

At maximum load in 3D, the core temperature reached 52 °C. In this mode, the fans operated at speeds of up to 1846 rpm, which caused a noise level of up to 27.6 dBA. This noise level is rated as quiet. An audio recording of the sound of the fans is available here.

The noise level spectrogram looks like this:

MTT S70

In idle mode in 2D, the temperature did not exceed 22 °C, while the fans operated at 1000 rpm, providing a noise level of 22 dBA.

When watching a movie using hardware decoding, the parameters remained stable.

At maximum load in 3D mode, the core temperature reached 46 °C. In this mode, the fans operated at speeds of up to 1828 rpm, which caused a noise level of up to 27.4 dBA. This noise level is also characterized as quiet. An audio recording of the sound of the fans is available here.

The noise level spectrogram has the following form:

MTT S70

Apart from the traditional quick start guide and warranty card (both in Chinese), there is nothing else included in the package.

MTT S80

As we have already noted, this card came to us as part of the system unit, so there was nothing with it except a power adapter. However, as a bonus, a huge branded Moore Threads mouse pad arrived in a separate package.

List of testing tools

All gaming tests used low graphics quality settings.

• Ashes of the Singularity
• Dota 2
• PlayerUnknown's Battlegrounds
• Counter-Strike: Global Offensive
• Heroes of the Storm
• World of Tanks
• Monster Hunter: World
• Far Cry 5
• Crysis 3

Test results in 3D games at 1920x1080 resolution

There were only nine tests, but the preparation for testing was, believe me, long and painstaking: the games required were not that old, oriented towards DirectX 11 or DirectX 10, and both MTT video cards had to work stably in them. The list of officially supported games from the manufacturer’s website consists of 90% of some Chinese titles unknown on the world market; it makes no sense to list it.

At the same time, we had to test the closest (if possible) competitors in the same games, for which we took:

• Nvidia GeForce GTX 1650 4 GB (Palit GeForce GTX 1650 StormX)
• Nvidia GeForce RTX 3050 8 GB (Palit GeForce GTX 3050 StormX)
• AMD Radeon RX 6500 XT 4 GB (Gigabyte Radeon RX 6500 XT Gaming)
• Intel Arc A380 6 GB (Gigabyte Intel Arc A380 Gaming)

It is clear that the level of MTT cards is a resolution no higher than Full HD (rather, clearly lower). It is also clear that using the maximum graphics level for such experiments is pointless; even with average quality, games often stalled, so all tests were ultimately carried out on low graphics settings in 1080p resolution.

Moore Threads MTT S80/S70 in general

It is difficult to draw clear conclusions: the cards have a huge number of disadvantages, varying in size and importance, but there are also a small number of advantages, also different.

Definitely a plus is the very fact of the emergence of a company that wants to play in the same league with such monsters as Nvidia, AMD and Intel. At the same time, Moore Threads Technology is focused exclusively on domestic Chinese and its own technologies, following the general trend of Chinese policy: to achieve the maximum possible autonomy in all spheres of life.

Moore Threads Technology (MTT), although backed by the Chinese government, keeps its activities strictly secret, releasing only limited information to the public. Its products are available on the market, but much of MUSA's architecture and details remain hidden. Unlike AMD, Nvidia and Intel, MTT avoids communicating with the press and does not disclose all the details of its developments.

There are a lot of rumors about the company, including suggestions that actual management is different from what is stated, as well as speculation about GPU die manufacturing outside of China. However, one can only rely on facts: the company exists, its products are on the market.

The main problem with MTT video cards is the drivers. Unlike Intel's approach, MTT started with support for older games and is gradually adding support for new APIs. This is questionable, but it may be because MTT doesn't have the same experience working with game developers as Intel, which can make driver optimization difficult.

Competing successfully in the global market requires collaboration with leading game developers, something MTT may not yet be able to achieve due to political and other barriers. However, the fact that the company was able to create working video cards in a short period of time testifies to the professionalism of its team.

MTT, as a new competitor in the GPU market, may not be fully on par with the «big» manufacturers just yet, but its emergence is still important to create competition in this space.

As a result, MTT S80/S70 work as gaming video cards, allow you to run some games at a level no higher than DirectX 11 and provide decent comfort in resolutions no higher than Full HD with low graphics quality. It’s very difficult to discuss prices, because the cards are sold only in the domestic market of the People’s Republic of China, and we can only say that in six months prices have dropped significantly compared to the initial ones.

Moore Threads MTT S80 (16 GB)

Overall, the tests showed that the Moore Threads MTT S80 (16 GB) is significantly inferior to the Nvidia GeForce GTX 1650 and competes more with the GeForce GTX 1050 Ti or even the GeForce GTX 1050. Perhaps with improved drivers and debugging, the situation will change, given the impressive characteristics of MTT S80 on paper. However, for now the picture looks like this. It's also worth noting that DirectX 12 support for this card could be problematic if it is indeed based on the old Imagination Tech architecture, which has been redesigned for new APIs.

During research, it turned out that the card has excellent hardware ability to decode video streams, but there is no OpenCL support, which creates problems in video editing programs such as Adobe Premiere. This is again due to MTT driver issues.

Despite the quiet cooler, the MTT S80 graphics card has high power consumption for its performance. In addition, the GPU and memory operate at high frequencies even in idle mode, which does not reduce power consumption in 2D mode. Perhaps future driver versions will improve this situation. It would also be helpful to add a backlight switch so users can control the orange LED on the card.

Moore Threads MTT S70 (7 GB)

When it comes to the Moore Threads MTT S70 (7 GB), the approach to creating this lower version of the accelerator looks unusual. Instead of simply cutting the video memory capacity in half compared to the older variant, the company has also removed one of the eight memory chips, resulting in a total capacity of 7GB instead of 16GB. At the same time, the number of execution units on the card remained approximately the same.

Obviously, 16 GB is overkill for a gaming graphics card of the level of the MTT S80, and perhaps the 16 GB version has a more specific use than just gaming. However, the performance difference between the S80 and S70 is small and does not appear to be directly related to memory capacity. However, due to incompletely optimized drivers in most games, MTT accelerators do not operate at full capacity, making it difficult to draw general conclusions about performance. In practice, however, the S70 only performs marginally worse than the S80 if these games run at all on the MTT cards.

As for the characteristics of the video card itself, all previously discussed aspects related to the S80 are applicable to it.

At the moment, MTT video cards are not products for wide mass consumption — they require lengthy debugging and improvements. Buyers of these cards are forced to become, in essence, beta testers using their funds.

However, it is important to note a positive point: efforts to create their own technologies, independent of American developments. One of the main advantages of this approach is the ability to increase competition, especially in the context of the inflated demands of Californian GPU manufacturers, especially the companies with the green logo. Can this small company grow to become a major player in the 3D graphics industry? Time will show.

We promise this review is just the beginning. After six months, we will re-test and release an update or addition if enough information has been collected.