Jin Lei From Aofei Temple

Qubit | Public account QbitAI

He is coming, he is coming.

HandheldGPU, wearing ablack leather jacket, jeans bottoms, yellow He has long skin and long hair, and his tough-guy aura cannot be concealed under his black-framed glasses.

But he is not NVIDIA’s Lao Huang.

He is Lao Zhang, who was once one of Lao Huang’s right-hand men——

Zhang Jianzhong, James, previously served as the top leader of NVIDIA China for 15 years.

The more well-known title now is domestic GPU companyMoore Thread’s founder and CEO.

The GPU that has just been heated up this time is already mass-produced and launched this yearThe second——

Time interval, only Half a yearlong.

Not only that, there are more tags attached to this GPU:

Domestic、Full-featured、The world's first,Mid-high-end……

And based on it, Zhang Jianzhong even releasedChina's firstmid-to-high-endgame graphics card .

To be honest, it is hard to imagine that such "speed" was created by a company that was just established 2 years ago.

This inevitably raises questions:

While it is fast, can the quality and performance also keep up?

We might as well take a look together.

The world's first full-featured GPU supporting PCIe interface

Moore Thread's second domestic full-featured GPU, named"MT-Chunxiao"< /span>.

Zhang Jianzhong also released it as the first product.

It is understood that Chunxiao integrates 22 billion transistors, has a built-in 4096MUSA architecture general computing core and 128 tensor computing cores, and can support FP32, FP16 and INT8 and other calculation precision.

Other key parameters are as follows:

• GPU core frequency: 1.8GHz
• FP32 computing power: 14.4 TFLOPS
• INT8 computing power: 57.6 TOPS
• Video memory bandwidth: 448GB/s
• Video memory type: GDDR6

Zhang Jianzhong also mentioned at the scene that Chunxiao had unlocked a "world's first":

Because it is the only GPU supporting PCIe Gen5 interface.

(Many manufacturers A consensus has been reached that PCIe Gen5 will be the key development direction of consumer and enterprise storage devices in the future.)

So compared with the "Sudi" GPU released by Moore Threads half a year ago, what are the differences between the two?

Zhang Jianzhong said that Chunxiao has achieved comprehensive upgrades in the four major Moore thread GPU engines:

• The performance of modern graphics rendering engines can be improved by up to 3-5 times
• The performance of AI computing acceleration engines can be improved by up to 4 times
• The performance of the intelligent multimedia engine is improved by up to 4 times
• The performance of the physical simulation engine is improved by up to 2.5 times

Our first full Functional GPU Sudi is actually a mid-to-low-end processor that can meet the needs of domestic domestic applications in the GPU industry.

But for most mainstream users, they still expect higher performance GPUs, so We quickly released Chunxiao to satisfy high-end gamers and meet the graphics and computing needs of more users.

In this way, our products can cover all users in the high, middle and low end.

When it comes to games, Zhang Jianzhong released another "domestic first" product based on Chunxiao GPU.

China's first gaming graphics card

In fact, Moore Thread also released the graphics card product MTT S60 based on Sudi half a year ago.

But the "purpose" of this graphics card seems to be more industry-oriented, that is, the B-side.

This graphics card is based on ChunxiaoMTT S80 is the kind that can be touched by more people——China’s first gaming graphics card.

At the scene, Zhang Jianzhong also used an interesting word to describe it:"National Trend".

From a performance perspective, its 4096 programmable MUSA cores can provide 14.4TFLOPS per unit at a main frequency of 1.8GHz. Precision floating point computing power.

Similar to Chunxiao, MTT S80 is also the industry's first graphics card product equipped with a PCIe Gen5 interface:

With 16GB GDDR6 large capacity High-speed video memory, coupled with 8K ultra-high definition and 1080P 360Hz high refresh rate display output capabilities, can bring a great experience to gamers.

It seems that "all talk but no practice" is not the style of Moore's thread conference. Just like last time, Zhang Jianzhong was also at the scene Directly apply the effect.

For example, it has been adapted in the Windows environment"Diablo 3", and this game still requires relatively high graphics card performance.

With the blessing of MTT S80, even if the whole process is in 4K high-definition quality, the FPS can be maintained at around 60(The higher the FPS, the smoother the picture).

In addition, Zhang Jianzhong also showed the racing game enthusiasts' favorite"Need for Speed", the effect under MTT S80 can be said to be quite smooth:

It is understood that MTT S80 has built-in MUSA DirectX Driver module in the Windows driver and has completed the adaptation to dozens of mainstream games.

More importantly, Zhang Jianzhong said that this graphics card will be available inDouble ElevenOn that dayLimited sales .

So we can look forward to the actual effect after getting it and whether the price is good.

New full-featured server GPU product

In terms of server products, Moore Thread has also been updated this time——MTT S3000.

Similarly, it is also based on the MUSA architecture and Chunxiao GPU, and its computing power can cover the complete MUSA software stack of graphics rendering, video processing, and deep learning.

Supported scenarios include AI inference and training, cloud gaming, cloud rendering, video cloud, digital twins, digital content creation, etc.

From a performance perspective, MTT S3000 includes 4096 MUSA stream processing cores and 128 dedicated tensor computing cores, with a transistor scale of 22 billion.

Its operating frequency is 1.9GHz, the video memory width is 256bit; with 32GB GDDR6 video memory, the bandwidth is 448GB/s; it can support FP32, FP16, INT8 and other calculation accuracy, among which the FP32 computing power can reach 15.2TFLOPS.

In Zhang Jianzhong’s opinion,Ecological collaboration is crucial to the advancement of AI applications.

Therefore, MTT S3000 is also compatible with PyTorch, TensorFlow, and Baidu Flying Paddle（PaddlePaddle）, plan map（Jittor） and many more It is a mainstream deep learning framework and realizes the acceleration of dozens of AI models such as Transformer, CNN, and RNN.

And MTT S3000 can be said to have "advanced with the times". The recently popular AI painting: Disco Diffusion and Stable Diffusion can also hold live.

Not just a hardware update

In addition to the above-mentioned hardware products, looking at the normal press conference, "combination of software and hardware"It is also a major feature of Moore's thread.

This is actually not difficult to understand. During our communication with Zhang Jianzhong, he also revealed the reason:

The software ecosystem is to promote GPU Key to the spread of computing.

GPU R&D system is very complex. Only hardware can be used to develop software. After the software runs on the hardware, more optimization needs to be done; after optimization, flaws in the architecture will be discovered, and in turn the hardware must be optimized.

Hardware and software are a process of mutual iteration and continuous improvement.

To this end, Moore Thread uses the MUSA architecture as the core this time, and Moore Thread has released a completeMUSA software stack.

The purpose is to serve the majority of developers and end users.

In addition, Moore Thread has made corresponding new actions in GPU cloud native, metaverse, and AIGC.

One ​​More Thing

Still focused In Zhang Jianzhong’s outfit this time:

This leather jacket, well, is really interesting.

— End —

Qubit QbitAI · Toutiao Signing Agreement

Follow us and learn about cutting-edge technology trends as soon as possible