Reporting by XinZhiyuan
Editor: Yuan Xie Good sleepy David
【New Zhiyuan Guide】China's domestic GPU is about to enter the market? Moore thread released a general-purpose GPU chip and 2 graphics cards, and Biling Technology also announced that its general-purpose GPU chip BR100 was successfully lit up once.
In the past two days, the domestic chip industry is full of "big news"!
On the 30th, The Moore thread led by Zhang Jianzhong, the former head of NVIDIA China, launched the domestic general-purpose GPU chip "Su Di", the desktop graphics card MTT S60 and the server graphics card MTT S2000.
On the 31st, Mr. Zhang Wen, a Harvard doctor and a senior investor in the investment industry, announced that his first general-purpose GPU chip BR100 series was successfully lit up on the same night.
The graphics card is sent twice in a row
When you think of a graphics card, what color can you think of?
It's nothing more than AMD's red and Nvidia's green, and maybe the black of the old yellow leather coat. By the way, these two days Intel also announced the launch of its own discrete graphics, so it may be necessary to add a touch of blue.
The field of graphics cards, to the larger extent, can affect the computing power of millions of AI laboratories in the country, and to a smaller extent, can change the game taste of millions of otaku.
This field is also looking forward to the rise of real domestic production day and night, and it has been extinguished by reality countless times.
So, what is in front of us this time?
Without further ado, go up the parameters first.
Benchmarking "International Flagship": 7nm process, parameters unknown
At about 11:00 p.m. on March 31, 2022, Biling Technology announced that the first general-purpose GPU chip BR100 series was successfully lit up at one time.
Bi Ling said that in terms of core performance design standards, BR100 is the largest general-purpose GPU chip in China, directly benchmarking the latest flagship products recently released by international manufacturers.
Based on its previous self-introduction on various platforms and the official announcement of this time, the BR100 series is based on the independent original chip architecture development of Biling Technology, using a 7nm process process, and combining cutting-edge design, manufacturing and packaging technologies including Chiplet.
Previously, on October 8, 2021, Biling Technology also announced that the general-purpose GPU chip BR100 was officially delivered to TSMC, began the tape-out process, and is expected to be released to the market in 2022.
Li Xinrong, former global vice president of AMD and current co-CEO of Biling Technology, said at his company's delivery tape-out celebration that this means that Biling Technology has entered the next stage.
"This represents that Biling Technology has truly entered the stage of talking about products, performance and achievements from a startup that talks about vision, team and technology."
Li Xinrong, former global vice president of AMD and current co-CEO of Biling Technology
When the tape-out was delivered, Li Xinrong said in an interview with Titanium Media that the BR100 was benchmarked against NVIDIA's 5nm GPU chip that was in the pipeline at the time.
Li Xinrong's introduction to Titanium Media is: "This chip adopts TSMC's 7nm process process, integrates PCIE Gen5 and HBM2e, etc. A combination of technologies such as 3D stacking technology is used."
36Kr's report on this matter has a sentence: "From the parameter point of view, if this chip can be successfully listed, it is also the AI chip with the largest computing power and the largest area at present."
However, looking at the reports of the two events on the whole network, whether the interviewer or the company disclosed itself, in addition to the general description of the "three advantages of high computing power, high versatility and high energy efficiency" and the above information. The specific performance, specifications and other parameters of the BR100 have not yet been revealed.
Considering that the chip is still in the "smooth progress of testing and tuning", it seems reasonable not to disclose specific parameters.
Zhang Wen: Yamanaka epiphany, to be a "Chinese core"?
The details of the entrepreneurial story of Biling Technology are basically released by the founder Zhang Wen in various interviews. I have to say that the story is more vivid, both like the part of the grid depiction, but also the creativity of the side oblique.
In an interview with Chuangyebang and The Paper, Mr. Zhang Wen said that it was not easy but experienced to start a business.
Zhang Wen, founder, chairman and CEO of Biling Technology
In the 1990s, Zhang Wen was a Ph.D. student at Harvard University, "only able to sleep three or four hours a day, reading obscure materials, this painful time, he became more and more resilient."
Resilience is difficult to quantify, but Mr. Zhang Wen said later in an interview with Chuang Chuang Bang that entrepreneurship has "greatly compressed its sleep time".
It seems that workers must be wary of bosses who have read a Ph.D., and people have the skills of not sleeping.
When he founded Biling Technology in 2019, according to Mr. Zhang Wen, "As a serial entrepreneur, this is the most prepared of all my entrepreneurship." "This is my last venture, the last company, and I decided to do a good job."
Then, in November 2021, Zhang Wen started another business and opened an L4 autonomous driving company "Yunji Zhixing".
Of course, the cliché part of the story cannot lack the anti-angles that are eventually indoctrinated. "A famous investor said, 'As long as you don't make chips, I'll invest in you.'" Zhang Wen's answer clearly disappointed the investor.
Even so, investors are still unwilling to give up such a high-quality entrepreneur and decide to invite Zhang Wen to meet and talk in detail. In just one hour, Zhang Wen... This investor was deeply impressed, and an investment was finalized."
In an exclusive interview with The Paper, Mr. Zhang Wen said that "China Core" is the most core component of the entire industry of Chinese Intelligent. "I started my business with the mentality of only taking a salary for one dollar, hoping to do a good job on this track."
Considering that the personal income tax is different from the capital gains tax system in China and the United States, Mr. Zhang Wen's statement may really be different from bezos. In China, if you don't take a salary and only take equity, maybe it is not really for tax avoidance, you can't avoid it.
In addition to the products that have not yet been sold, the most proud thing for the founders should be the talent lineup of Jijiduoshi.
In an exclusive interview with Chuang Chuang Bang and The Paper, Mr. Zhang Wen repeatedly mentioned his achievements in recruiting talents: "The best semiconductor talents in the Chinese are concentrated, and many of our team have been doing several generations of GPUs." "The 'Dream Team' of semiconductor talents, from technology to engineering to business, the team includes many top scientists, as well as industrial talents from the world's leading chip companies with more than 20 years of experience."
According to the press release of the 2021 product streaming celebration, Biling Technology has nearly 600 team members in Shanghai, Beijing, Zhuhai, Hangzhou, Chengdu, and the United States.
In mr. Zhang Wen's self-narrated entrepreneurial story, there is an interesting little detail.
In an exclusive interview with Chuang Chuang Bang, Zhang Wen said that he traveled to Wuyi Mountain, "a huge mountain with four big characters engraved on it as 'Bi Li Wan Ling' stands in front of him. This mountain is the largest naturally formed rock in Asia."
Zhang Wen: "After sitting quietly at Zhu Xi Academy in the mountains for a whole afternoon, I secretly made up my mind to become a company as tall and rock-solid as this boulder.
Although the bridge section of Yamanaka's epiphany is cheesy, the appearance of this company name is interesting.
"Bili Wanling" is a play of the uncultured operators of tourist attractions, and this idiom usually appears in Chinese as "Bili Qianling". After all, seven feet is a slur, the mountain is seven thousand feet high, and the mountain height of seven thousand feet is mysterious.
The most well-known use of "Bi Li Qian Ling" is the legendary Lin Zexu's collection of sentence couplets that refused to bribe: "Hai Na Bai Chuan, there is tolerance is great." The wall stands on a thousand feet, and if there is no desire, it is just." In case everyone is embarrassed, he gently asks potential stakeholders not to send too much money to invest in his career.
As for the origin of the Chinese of "Bili Qianling" in the 3rd and 6th centuries AD, it is used to describe the nature of the area of "gusa town", "building but enemy", and "throwing foot into nothing" and other natural dangers such as Sword Pavilion, QiuChi, and Onion Ridge.
These allusions seem to be "incompatible" with the record amount of financing?
The story can be played, and the numbers are more real: bi-ling technology has been surrounded by the aura of money since its birth.
On September 9, 2019, Biling Technology was registered and established. On June 16, 2020, Biling Technology announced the completion of a series A financing with a total of 1.1 billion yuan, setting a record for the scale of Series A financing in China's chip design industry at that time.
According to its own commitment, from its incorporation to the completion of the B round of financing in March 2021, Biling Technology has raised more than 4.7 billion yuan in 18 months, setting a financing record for Chinese chip startups.
According to the statistics of Jiazi Lightyear, from 2020 to 2021, there are nearly 20 financing incidents in the field of GPGPU, of which four are the financing of Biling Technology.
At the end of March 2022, Biling Technology released the news that it would continue to raise funds: the company set its own valuation at 17 billion yuan and hoped to raise up to 2 billion yuan. Discussions are still ongoing and funding targets and valuations are subject to change.
It took 14 months and took the lead in mass production of 12nm graphics cards
On November 25, 2021, Moore Thread announced that it had completed the development of its first domestic full-featured GPU in less than 300 days.
After 4 months, Moore Thread officially released the desktop graphics MTT S60 and server graphics MTT S2000 on March 30, 2022.
From the paper parameters, the MTT S60 has:
12nm process
2048 MUSA cores
8GB LPDDR4x video memory
Single-precision floating-point arithmetic 6 TFLOPS
The pixel fill rate is 192G Pixel/s
Supports up to 8K
In contrast, the launch of the NVIDIA GTX 1070 in 2016 has achieved 6.5 TFLOPS of computing power, and the Xbox One X in 2017 also reached 6 TFLOPS.
It is said that the MTT S60 can run "League of Legends" smoothly at the highest image quality at 1080p resolution, and it is not known how much the frame rate can be stabilized.
Then again, if you want to run smoothly, you only need intel UHD 630 episode display. On their own, the NVIDIA GeForce 560 and AMD Radeon HD 6950 are also available. However, the so-called "smooth" now seems to have a tendency to move towards 120 frames.
It is worth mentioning that MTT S60 supports mainstream CPUs such as Intel, AMD, Loongson, Feiteng, and Mega core, as well as operating systems such as Windows 10, Kirin, Tongxin, and Ubuntu.
In order to match the multi-GPU configuration mode of the data center, the MTT S2000 adopts a single-slot passive heat dissipation design. Its paper parameters are:
4096 MUSA cores
Up to 32GB of video memory
Single-precision floating-point arithmetic 12 TFLOPS
In the AI model algorithm, it can accelerate deep learning frameworks such as Tensorflow and Pytorch, and support CPUs with x86 and ARM architectures.
As you may have noticed, what does "MUSA" mean in the argument?
MUSA (Moore Threads Unified System Architecture), also known as the "unified system architecture" of Moore threads.
These include a unified programming model, software runtimes, driver frameworks, instruction set architectures, and chip architectures.
Applications developed on MUSA with wide portability can run on numerous computing platforms in the cloud and at the edge.
The first generation of full-featured GPU chips based on MUSA - Sudi.
According to reports, this so-called "full-featured" refers to the fact that "Su Causeway" has four mainstream engines built-in:
Graphics rendering engine
Intelligent multimedia engine
AI compute acceleration engine
Physics simulation and scientific computing engine
Well, judging from the GIF of the qubit transmitted back at the first scene, it should be quite good.
4.9MB
High-precision texture rendering
4.0MB
Global illumination and ambient occlusion
In addition, AlphaCore, a next-generation multi-platform GPU physics simulation system independently designed and developed by Moore Thread, can perform ultra-high-precision physical simulation processing of complex solids, fluids and other effects in the physical world.
The most powerful thing is that AlphaCore is compatible with runtime API environments such as Vulkan, CUDA, and DirectX in the existing ecosystem.
In this way, it is "more than a star and a half stronger" than AMD, which does not support CUDA.
Zhang Jianzhong: 19 months away from NVIDIA to get a domestic GPU
Zhang Jianzhong, the founder and CEO of Moore Thread, graduated from the Department of Computer Science of Nanjing University of Science and Technology, and then completed his master's degree at the Institute of Automation of the Ministry of Metallurgy. He has been employed by Dell, HP and other enterprises, serving as sales director and general manager of business units.
Jianzhong Zhang joined NVIDIA in May 2005 and eventually served as NVIDIA's Global Vice President and General Manager for China. He left Nvidia in the fall of 2020. Moore Threads was registered in October 2020.
During his tenure at NVIDIA, NVIDIA's GPU market share in China was less than 50% in 2008 and more than 80% in 2020. Impressive performance.
In a February 2021 funding press release, Moore Thread claimed to have "GPU ecosystem and industry marketing experts who helped Nvidia succeed in China", which is really not a fool's errand.
Resources:
https://tech.sina.cn/2020-06-16/detail-iirczymk7222495.d.html
https://www.tomshardware.com/news/first-wholly-domestic-chinese-GPU-graphics-card
https://www.birentech.com/news/165.html
https://www.birentech.com/news/128.html
https://www.36kr.com/p/1433660203105928
https://www.birentech.com/news/119.html
https://www.huxiu.com/article/493306.html
shorturl.at/adCKX