Gráfica GPU made in China

[Dark Kaeser]

Já há alguns modelos de GPU chineses, assim fica um tópico dedicado à coisa.

- Innosilicon

Linha Fantasy (ou Fenghua em chinês), que até têm um licenciamento da Imagination (não sei se toda a linha é ou não licenciada)

Fantasy One A

a Fantasy B, é um MCM de 2 chips da placa acima

https://www.innosilicon.com/html/gpu/#a1


- Jingjia Micro

Depois do anúncio há coisa de um ano do tape out de 2 GPU da série JM9, para o mercado servidor pois as mesmas nem suportavam DirectX, nunca mais vi ou ouvi nada

https://videocardz.com/newz/jingjia...-to-gtx-1080-performance-officially-tapes-out


- Glenfly

Parece ter apenas um modelos dGPU o Arise GT10C0, não percebi se a empresa está afiliada à Zhaoxin (cpu x86)

https://www.glenfly.com/index.php?catid=22


- Biren

Tem o Biren BR100, GPU apenas para servidores, no formato OAM ou PCIe

https://www.birentech.com/product_details/1005557637772464128.html

https://www.birentech.com/product_details/1005557844745474048.html

https://www.nextplatform.com/2022/08/25/china-launches-the-inevitable-indigenous-gpu/


- Moore Threads

Já tinham apresentado alguns modelos para desktop, workstation e servidores no início do ano, mas agora apresentam os "topo de gama"

Moore Threads MTT S80

https://www.mthreads.com/product/S80

MTT S80 uses a 12NM process, 4096 MUSA cores, and a maximum video memory configuration of 32GB. The maximum single-precision computing power is 14.4TFlops. It can encode and decode H.264, H.265, AV1 multi-channel HD video and perform a variety of AI model calculations. compel acceleration To meet the high-density GPU configuration in the data center, the MTT S80 uses passive cooling and a single-slot design.
Moore Threads has released a unified programming model, runtime library, driver, and other software tools for the MTT S80 series products' hardware architecture, which can help developers complete application porting and adaptation and fully call MTT OpenGL ES, DirectX, Vulkan, and other graphics APIs. Support the audio and video processing ecosystems with FFMPEG and VA-API/DXVA compatibility, and meet AI and scientific computing program compatibility with OpenCL and Vulkan.
Moore Threads MTT S80 is compatible with X86, ARM, and other CPU architectures, as well as mainstream Linux operating system distributions, and has begun to collaborate with many server partners to adapt to a wide range of general-purpose and GPU server models from OEM manufacturers. In the application environment, the deployment is complete.

https://www.unixcloud.ltd/product/mtt-s80/

Moore Threads S3000

https://www.mthreads.com/product/S3000

MTT S3000 employs a 12NM process, 4096 MUSA cores, single-precision computing power of up to 15.2TFlops, 32GB video memory, supports H.264, H.265, AV1 multi-channel high-definition video encoding, and multi-channel AV1, H.264, H.265, VP9, VP8, MPEG1/2/4, AVS2 and other formats of ultra-high-definition video decoding, and can support MTT S3000 is a single-slot, half-height, half-length PCIE card optimized for video scenarios. It uses passive cooling and has a maximum power consumption of 35W.
Moore Threads incorporates Moore Thread's first-generation intelligent multimedia engine, MTSmart Media Engine for MTT S3000, which can accelerate a variety of audio and video applications on the cloud and edge sides, including security intelligence, video conferencing, live on-demand, cloud gaming, cloud desktop, and so on. Supports the audio and video processing ecosystem by being compatible with the FFmpegy audio and video framework and the VA-API application program interface, providing customers with the underlying GPU hardware encoding and decoding capabilities; supports OpenGL, OpenGL ES, Vulkan, and other graphics APIs; and is compatible with Pytorch, TensorFlow, PaddlePaddle, and other frameworks. A variety of artificial intelligence computing frameworks; through OpenCL to meet AI and scientific computing program compatibility.
Moore Threads MTT S3000 is backwards compatible with X86, ARM, and other CPU architectures, as well as popular Linux operating systems.

https://www.unixcloud.ltd/product/mtt-s3000/

​

 

[Dark Kaeser]

MTT S80 Chinese gaming GPU goes on sale, but you can’t buy it without a motherboard​

Over at Chinese retail website, you will find the first listing of the Moore Threads MTT S80. The card is available for 2999 RMB, but this is not the price of the graphics card itself.

This card is currently only available in a bundle with ASUS TUF GAMING B660M-PLUS D4 worth 1029 RMB. This means that the value of the M80 GPU is around 1970 RMB (276 USD). Unfortunately, there is no option to buy it separately.

https://videocardz.com/newz/mtt-s80...ale-but-you-cant-buy-it-without-a-motherboard

 

[Dark Kaeser]

1os testes da MTT 80

MTT S80 supports both DirectX9 and DirectX11 APIs. However, the software is still under active development, so it does not support DX11 fully yet. For this reason, conventional tests of S80 are not possible.

In actual game tests the card performs fine at 1080 low settings. We are talking about 213 FPS in CSGO, around 144 FPS in League of Legends, 180 FPS in Crossfire and 90-100 FPS in Diablo 3. Some reviewers tested games such as Minecraft, but the GPU has trouble working with the Windows version, but it is otherwise playable at 40 to 50 FPS.

Power consumption of this card in standby mode is 114.6W, a truly unacceptable figure from any GPU in 2022.

https://videocardz.com/newz/moore-t...orce-gtx-1060-performance-and-110w-idle-power

 

[Dark Kaeser]

Review deve estar para breve....

https://twitter.com/i/web/status/1620769446947332096

 

[Dark Kaeser]

https://twitter.com/i/web/status/1620827468449062916

https://twitter.com/i/web/status/1620871294257598465

 

[Dark Kaeser]

Chinese GPU Targets GeForce GTX 1650 Performance​

Zhihui's IDM929 graphics processor is based on the company's proprietary IDMV architecture, but the company does not disclose its feature set. The only things we know is that the GPU has a pixel fill rate of 19.2 GPixels/s, a texture fill rate of 76.8 GTexel/s and compute performance of 2.5 TFLOPS (presumably FP32), which is comparable to that of Nvidia's GeForce GTX 1650 (which was released in mid-2020). The graphics processor also supports video decoding in H.264, MPEG2, MPEG4, VC-1, DivX, and VP6 formats. Meanwhile, the unit has four display engines and supports DVI, HDMI, and D-Sub/VGA interfaces.

The GPU is made using a 14nm-class process technology, which could possibly mean that it is made by Semiconductor Manufacturing International Corp. (SMIC). It has a clock frequency of up to 1.2 GHz, which is a relatively low frequency for a chip made on a 14nm node.

Zhihui Microelectronics is a new name in the rapidly developing Chinese GPU industry, but its main investor Hangzhou Guoxin Technology is a relatively well-known developer of embedded system-on-chips for applications like TVs. That said, it is possible that Zhihui will target primarily embedded applications, at least initially. Meanwhile, Zhihui and Guoxin intend to offer a platform consisting of a CPU and GPU, though it is not directly said whether it will be aimed at PCs or embedded applications.

https://www.tomshardware.com/news/c...tm_source=twitter.com&utm_campaign=socialflow

 

[Dark Kaeser]

W..t...

The company has just announced the MTT S70 desktop GPU featuring cut-down specs compared to the S80 released a few months ago. Unfortunately, MTT graphics are not powerful enough to compete with anything AMD, NVIDIA, and Intel currently offer, that’s despite rather impressive specs on paper.

The S70 features 3584 MUSA cores and has a boost clock up to 1.6 GHz. This is a downgrade from 4096 cores and 1.8 GHz boost on the S80. As a result, the S70 offers 3.2 TFLOPs (22%) lower compute performance.

The card offers an unusual 7GB of GDDR6 memory offering 392 GB/s of maximum bandwidth. This means that the GPU has a 224-bit interface which is not very common on graphics cards.

https://videocardz.com/newz/moore-t...-with-3584-cores-and-unusual-7gb-gddr6-memory

 

[ZIGS]

Chinesices lol

 

[ímpar]

Chinese Powerstar x86 CPUs Were Made In Partnership With Intel, Use Custom Design
...
Following this controversial reveal, Powerleader has now published an official statement in which the company states that the new CPUs were designed in partnership with Intel and utilize a slightly custom design. The company doesn't go into a lot of detail as to what these custom changes are but it could be related to software and firmware-level support to make them appear as Powerstar chips instead of the Intel Core CPUs that they really are.
...
https://wccftech.com/chinese-powers...-in-partnership-with-intel-use-custom-design/

 

[Dark Kaeser]

Chinese company Metax reveals N100 AI/video processor, gaming GPUs planned for 2025​

The company has showcased its N100 accelerator at the “China International Social Public Safety Products Expo”. Attendees were also given a first look at the MXN100 GPU covered with a large integrated heat spreader. Underneath there should be a HBM2e memory of unknown capacity.

The main feature of this processor is video processing of 1 channel encoding and up to 96 decoding in HEVC, AV1, AVS2 and H.264 formats at the same time and up to 8K resolution.

Capable video processing power would be combined with AI inference processing up to 160 TOPS (trillion operations per second). The N100 could serve in intelligent video processing for transportation and cloud computing. It is unlikely to compete with AMD Instinct or NVIDIA Data-Center GPU series.

https://videocardz.com/newz/chinese...-video-processor-gaming-gpus-planned-for-2025

 

[Dark Kaeser]

https://twitter.com/i/web/status/1668567207121502210

https://twitter.com/i/web/status/1668635869245472771

modernices...

 

[ZIGS]

Mas corre o Crysis?

 

[Dark Kaeser]

Moore Threads latest GPU driver enables Open GL 3.3 support, up to 40% improvement for CryEngine and 20 fixes

One standout improvement is the introduction of OpenGL 3.3 support. Even though this API is relatively old, it lacked compatibility with MTT GPUs before

https://videocardz.com/newz/moore-t...-to-40-improvement-for-cryengine-and-20-fixes

 

[ímpar]

Chinese GPU Maker, GITSTAR, Intros JH920 Graphics Card: Faster Than GTX 1050 With AMD FSR Support

https://wccftech.com/chinese-gpu-gi...rd-faster-than-gtx-1050-with-amd-fsr-support/

 

[FastBoy]

Chinesices lol

Não substimes, isso fora do mercado Europeu e Americano é capaz de expandir largamente em termos de volume... como os carros, quando deres por ela, estará a fazer pressão na industria Americana e Europeia sem se dar conta.

 

[ímpar]

Moore Threads Prepares S90 and S4000 GPUs for Gaming and Data Center

https://www.techpowerup.com/314908/...s90-and-s4000-gpus-for-gaming-and-data-center

 

[Miguel_Pereira]

Isto é fabricado em Taiwan ou a China já tem foundries?

 

[Dark Kaeser]

A China sempre teve foundries, a mais avançada é a SMIC, que devido às restrições impostas pelos EUA não tem, neste momento, acesso às mais avançadas fontes EUV, da neerlandesa ASML, que permitem reduzir uma série de etapas na produção de chips em litografias mais avançadas, pelo que neste momento o mais avançado de que dispõem é equiparado N7 da TSMC, segundo os dados da TechInsight.

SMIC process is 7nm, says Tech Insights​

https://www.eenewseurope.com/en/smic-process-is-7nm-says-tech-insights/

Huawei planning on using SMIC to fab 7nm ICs this year​

The 7nm process Huawei is using is thought to be SMIC’s N+1 process characterised by TechInsights as having Fin Pitch (FP), Contacted Poly Pitch (CPP) and Metal 2 Pitch (M2P) sizes either larger or the same as TSMC’s N10 process.

However Extensive Design Technology Co-Optimization (DTCO) features and high-density logic libraries enable a logic transistor density of 89 million transistors per square millimeter (89MT/mm^2), which is comparable to TSMC’s N7 and Intel’s 10nm, which makes SMIC’s N+1 process a viable 7nm-class alternative for logicng SMziC to fab 7nm ICs this year

https://www.electronicsweekly.com/n...-using-smic-to-fab-7nm-ics-this-year-2023-07/

China Accelerates Foundry, Power Semi Efforts​

• CR Micro — Chongqing (300mm, legacy nodes)
• Hua Hong — Wuxi (300mm, legacy nodes)
• Nexchip — Hefei (300mm, legacy nodes)
• SMIC — Shanghai (300mm, 14nm and above)
• SMIC — Beijing (300mm, 28nm and up)
• Wingtech — Shanghai (300mm, legacy nodes)

Shanghai Huali has two fabs, which are full. The company is 28nm-capable with 14nm slated by year’s end.

https://semiengineering.com/china-accelerates-foundry-power-semi-efforts/

 

[Dark Kaeser]

Chinese MTT S80 PCIe 5.0 GPU Closes in on GTX 1650​

New driver optimizations have enabled Chinese vendor Moore Threads' MTT S80 gaming graphics card to rival Nvidia's GeForce GTX 1650 at 4K gaming.

The MTT S80 reportedly leverages the Chunxiao silicon, based on the PowerVR architecture developed by Imagination Technologies. The exact generation of the PowerVR GPU remains a mystery. Nonetheless, there's a lot of optimization work to get the graphics card to play nice with DirectX 11 titles.

https://www.tomshardware.com/news/chinese-mtt-s80-pcie-50-gpu-closes-in-on-gtx-1650

 

[ímpar]

China’s Loongson Plans To Tape-Out A GPU In Q3 2024 That Is Equivalent To AMD’s 6-Year Old RX 550

https://wccftech.com/china-loongson-tape-out-gpu-q3-2024-equivalent-to-amd-6-year-old-rx-550/