Chinese semiconductor industry

tphuang said:

National standard might be a good thing or it might not. I think more likely, one domestic player in the AI space will be the big winner and maybe a couple of smaller ones by each of the cloud service providers. The big winner looks like Biren technology right now based on the performance of their chips. I think it is more important that people who are familiar with CUDA can easily use the Chinese one. That way, companies outside of China will also be interested in using Chinese AI chips/servers. After all, Chinese companies will want to sell their GPUs abroad.

I took a look at CUDA just now and it looks very much just like C++. Looks like BirenSupa also has a c++ extension

Please, Log in or Register to view URLs content!

Huawei is an interesting case actually. I haven't seen them buying into any of these promising chip companies. Inspur/Baidu are buying into/working with both BR100 and Kunlun-2. I speculate Alibaba cloud will eventually purchase BR100 also. Tencent mobile is looking to build its own chips and also invested in Enflame. Alibaba cloud developed their own server chip and also invested in Vastai technology.

Huawei doesn't look to have invested in anyone. Prior to the sanctions, it had a very robust chip making business in server and AI with Hisilicon. In fact, Kunpeng 920 is still being used widely 3 years later. Not as advanced as Alibaba's Yitian-710, but pretty good for something designed 3 years ago. They also developed their own Ascend series of AI chips. So, I can only presume that they are working hard at designing a newer more powerful AI GPU with both 12 nm and 7 nm process with their own fab or with SMIC. In many ways, Huawei has been forced to trail blaze for rest of China's semiconductor business.

Click to expand...

There are definitely arguments for free-market and competition to sort out winners, and let the best win. I'm afraid that's probably not the best strategy for China now in this industry.

First of all, China is a laggard in this industry and Chinese companies are not pioneers in this industry. They are all trying to desperately catch up. There is simply no time to experiment with the different models and sort out the winning ecosystems and winners. The US has demonstrated the model and road forward.

Secondly, China's strength is in her large market and potential large-scale applications. Quickly getting to some kind of industry standard will help grow the market and form the positive feedback loop a lot more quickly. Also a more or less unified standard will also reduce the wasteful investment in marketing and sales of each company's own ecosystem, so they can more focus on performance and other technical criteria. This actually helps pick the winners quicker.

Lastly, Chinese companies are not particularly experienced or well-known in setting industry standards and collaborating among themselves. Yet this is the area China and Chinese companies need to learn quickly in order to develop alternative technology ecosystems to compete with the US, not just in GPU or even in semiconductor. I'm not talking about national standards as in imposed by government, which don't always work well, particularly in new or emerging industries. It's more about industry standards set and adopted by partners and competitors in an industry. In that case, it doesn't even have to be one standard. Companies can form competing standards. In this particular case, the benefits of domestic companies collaborate to develop the industry and expand the downstream applications far outweighs the those from free-for-all competition with its attendant cost and inefficiencies.

I just going to say this but any person who wants to "expose as a fraud" a well known qualified scientist or distinguish member of the tech industry, people with heavy resumes, highly cited papers and so on. Those persons has to show off some qualifications themselves, good theoretical analysis and bring some repeatable experiments to the table, no write vindictive crap in a blog.

weig2000 said:

There are definitely arguments for free-market and competition to sort out winners, and let the best win. I'm afraid that's probably not the best strategy for China now in this industry.

First of all, China is a laggard in this industry and Chinese companies are not pioneers in this industry. They are all trying to desperately catch up. There is simply no time to experiment with the different models and sort out the winning ecosystems and winners. The US has demonstrated the model and road forward.

Secondly, China's strength is in her large market and potential large-scale applications. Quickly getting to some kind of industry standard will help grow the market and form the positive feedback loop a lot more quickly. Also a more or less unified standard will also reduce the wasteful investment in marketing and sales of each company's own ecosystem, so they can more focus on performance and other technical criteria. This actually helps pick the winners quicker.

Lastly, Chinese companies are not particularly experienced or well-known in setting industry standards and collaborating among themselves. Yet this is the area China and Chinese companies need to learn quickly in order to develop alternative technology ecosystems to compete with the US, not just in GPU or even in semiconductor. I'm not talking about national standards as in imposed by government, which don't always work well, particularly in new or emerging industries. It's more about industry standards set and adopted by partners and competitors in an industry. In that case, it doesn't even have to be one standard. Companies can form competing standards. In this particular case, the benefits of domestic companies collaborate to develop the industry and expand the downstream applications far outweighs the those from free-for-all competition with its attendant cost and inefficiencies.

Click to expand...

Well, I don't have any experience developing software in AI, but it seems to me that most people in the market are familiar with C++ and CUDA implementation right now. There is probably a set to features that anyone would expect. Based on the amount of time that Biren leadership have had in this industry, it looks like they have put together something that is very similar to CUDA. Keep in mind that AMD AI accelerator does not work with CUDA either. They use the open source openCL. Nvidia supports openCL and CUDA but are more optimized for the latter. AMD just supports openCL (it looks like). So, for the Chinese chipmakers, it seem most logical to just support openCL + their own proprietary one that's close in features to CUDA. The standard they seem to use is just knowledge of C++. I would love for someone to correct me here, but both openCL and CUDA seem to just be extensions over C++. This a documentation explaining openCL. It looks like just something that allows you to better utilize instruction sets of the GPU. From what I've seen with CUDA, it's the same thing.

It looks like Biren was pretty in touch with what you need to get people to jump on board to switch to their chips. They seem to support OpenCL as well as all the standard AI framework that people would use.

Developers who are familiar with (Nvidia’s) CUDA can easily write code for SUPA,” said Hong. Supported AI frameworks include PyTorch, TensorFlow and PaddlePaddle. The company also provides the OpenCL compiler. The dual-die BR100 appears as one GPU to the software layer.

Click to expand...

The more I look at this, the more I think it should be fairly easy for any self respecting AI developer to switch between AI software models/layers. I think the bigger issue for Chinese chip designer is that the customers will really work with them to have everything debugged and have frequent firmware updates. I'm not actually clear what Huawei is talking about here? Are they concerned that Biren, Baidu and Inspur will take over the market as the first mover and leave everyone else out? In the end of the day, you can have all the chip designers support openCL and a Chinese version of that, but I don't think it will ever be as optimized as the native "software model".

gelgoog said:

About China accepting the acquisition of Xilinx by AMD. I think this was a good call since after Altera was bought by Intel this is the only way to prevent Intel from concentrating way too much IP. From an anti-monopoly perspective it makes sense to allow this. And all these companies are based in the US to begin with. So it is not like this will make the US get more IP.

Click to expand...

paiemon said:

Bingo, AMD owning Xilinx or Xilinx standalone makes no difference to the product and its availability in China since its US content would cause it to fall under any export controls to say Huawei regardless of ownership. Its why the Nvidia takeover of ARM was dead on arrival, since it would result in one of the few non-US sources of semiconductor IP/content falling under the US umbrella. Even though ARM has US offices contributing to the content level, the core of its portfolio remains UK/OUS based meaning it is much harder to restrict with export controls.

Click to expand...

If AMD owning Xilinx and Xilinx standalone makes no difference to the product and its availability in China, then why should China approve of the merger. You guys seem to take the view that Anti-monopoly is a Chinese issue and not an American issue. If Intel's Altera acquisition resulted in a huge IP advantage (it has not), the people who lose out the most are other US companies (i.e., AMD, Xilinx), not China.

If Intel really was able to abuse its monopoly position over CPU/FPGA/accelerator IP then either FPGA competitors go out of business, turn to China for funding/business or force US gov/Intel to subsidize their existence. All of these scenarios are much worse for US than China.

this is how great the hearten server is (the one Biren is working on with Inspur)

徐凌杰表示，“用国际巨头在售的旗舰产品，6000台服务器（DGX A100 640GB？）可以达到15 EFLOPS的浮点算力，需要3000台机柜，占地空间1万平方米以上，峰值功耗39兆瓦，最高需要3.4亿度电每年，相当于4.2万吨煤的发电量。”

Click to expand...

而如果换成OAM服务器海玄，“只需要2000台海玄服务器，达到16 EFLOPS的算力；只需要1000个机柜，占地面积不超过3500平方米，峰值功耗14兆瓦，1.2亿度年用电量，相当于1.5万吨标准煤发电量，实现了整体方案64%的成本下降。”

Click to expand...

Basically, you can get the same calculation performance with 2000 Hearten as you would with 6000 DGX A100. It takes up 1/3 of the space and requires 36% of the energy to provide that much power.

the other thing I got a sense of is just how much the AI ecosystem is built around what Nvidia has to offer. It's great that the largest players in the market are working intimately with Biren and other Chinese chip designers. They need to be able to provide the resources needed to help people transition from Nvidia. Hence the creation of the Biren developer cloud already becoming active.

生态建立一直是壁仞科技发展战略的重中之重，为了更好地服务全球开发者，壁仞科技开发者云也已经正式上线，官网上已开放邀测。壁仞科技希望通过社会各界开发者的共同努力，形成聚沙成塔的力量，共同推动中国半导体技术的发展，真正做到让更多人从技术发展中受益。

Click to expand...

That's probably why Biren itself has done as many cooperation deals with so many Universities. talks here about training talent and working with them on various field like Medical imagine, EM simulation, molecular dynamics.

壁仞似乎和不少高校展开了合作，主要着力在和高校建立合作关系，从学术研究、人才培养和科研成果转换上入手。从壁仞展示的视频来看，目前涉及的合作方向包括医疗影像、分子动力学、电磁仿真等领域。虽然不清楚合作深度和成果产出如何，但这本身就是生态扩展、为未来打基础的长远方案。

Click to expand...

hmm from another article, this is interesting. Looks like Biren joined Baidu Paddle's creation of hardware ecosystem. Maybe Baidu PaddlePaddle is creating this for their machine learning framework and adding chip designers here and Huawei is not happy with it.

壁仞科技宣布加入由百度飞桨发起的硬件生态共创计划

Click to expand...

Ping An Technology will use Ping An cloud and work with Biren technology to create a high end general purposes intelligent computing platform. Biren will work to train Ping An technology on using their chips. Will help Ping An with machine learning and AI vision

平安科技董事长兼CEO黄宇翔表示，平安科技将以平安云为基础，结合壁仞科技产品共同打造高端通用智能的算力平台，针对不同的用户场景，推出有市场竞争力的产品和解决方案。壁仞科技还将依托国产大算力芯片为平安提供高效训练、低成本推理、综合应用硬件等解决方案，并且对平安在AI视觉、语音人机交互底层算法等业务场景进行软硬件专项适配。

Click to expand...

China Mobile joins forces with Biren also for their Jiutian AI platform .

中国移动研究院人工智能与智慧运营中心副总经理金镝表示，智能计算是整个算力网络发展中的先锋力量，希望壁仞科技的新产品与中国移动的算力网络整体的发展深度适配，同时也希望壁仞科技的产品和中国移动九天人工智能平台有更好的合作，为开发者和上层应用的使用者提供更加丰富的算法，更加强劲的算力和更加优秀的运营成本。

Click to expand...

Looks like another major customer is GDS Services

壁仞科技正在积极布局BR100商业化落地，已与平安科技、浪潮信息、万国数据等重要合作伙伴强强联手

Click to expand...

tokenanalyst said:

I just going to say this but any person who wants to "expose as a fraud" a well known qualified scientist or distinguish member of the tech industry, people with heavy resumes, highly cited papers and so on. Those persons has to show off some qualifications themselves, good theoretical analysis and bring some repeatable experiments to the table, no write vindictive crap in a blog.

Click to expand...

Exactly this. If you make huge claims about someone with proven results and credentials being a fraud - not overrated, not less important than imagined, but an outright fraud - then there is a very high standard for evidence.

A single blog post does not cut it for evidence.

For example, when anti Chinese racist professor Frank Dikotter wrote papers talking about how opium was not harmful to China, professors using their real names went on the record with highly cited papers in academic journals to refute him. They didn't post an anonymous blog.

FairAndUnbiased said:

Exactly this. If you make huge claims about someone with proven results and credentials being a fraud - not overrated, not less important than imagined, but an outright fraud - then there is a very high standard for evidence.

A single blog post does not cut it for evidence.

For example, when anti Chinese racist professor Frank Dikotter wrote papers talking about how opium was not harmful to China, professors using their real names went on the record with highly cited papers in academic journals to refute him. They didn't post an anonymous blog.

Click to expand...

I replied to his posts in the science and tech development thread. Technically quantum communication and quantum computing is still tangentially related to chinese semiconductor industry, but I think mostonhel's (misguided gripe) is just with Pan when it should be funding (in general, not just China) for quantum technologies in the first place. There are tons of critics (even within Academia) related to the hype and funding associated with quantum area of research, none of them are a China-specific problem.

tphuang said:

Keep in mind that AMD AI accelerator does not work with CUDA either. They use the open source openCL. Nvidia supports openCL and CUDA but are more optimized for the latter. AMD just supports openCL (it looks like). So, for the Chinese chipmakers, it seem most logical to just support openCL + their own proprietary one that's close in features to CUDA.

Click to expand...

OpenCL has been around for yonks. It has had its advantages but it is way behind CUDA as a development environment. The main issue is lack of 3rd party or even vendor made libraries made on top of it. OpenCL is basically compiled on application startup from source code kind of like a shader program, so you can't make binary libraries. Not unless you have OpenCL 1.4 or 2.0 anyway. Those can compile to byte code. CUDA has libraries on top of libraries on top of libraries of stuff you can use as an application developer. OpenCL environment is weak to say the least. Imagine having a C++ compiler but basically no libraries, only snippets of code you get from other people.

coolgod said:

If AMD owning Xilinx and Xilinx standalone makes no difference to the product and its availability in China, then why should China approve of the merger. You guys seem to take the view that Anti-monopoly is a Chinese issue and not an American issue. If Intel's Altera acquisition resulted in a huge IP advantage (it has not), the people who lose out the most are other US companies (i.e., AMD, Xilinx), not China.

If Intel really was able to abuse its monopoly position over CPU/FPGA/accelerator IP then either FPGA competitors go out of business, turn to China for funding/business or force US gov/Intel to subsidize their existence. All of these scenarios are much worse for US than China.

Click to expand...

China tried that with attempted acquisition of Lattice Semiconductor by Canyon Bridge Capital Partners. Donald Trump didn't allow it to happen. Lattice was like #3 in FPGAs and even then the sale did not happen because of US govt.

gelgoog said:

China tried that with attempted acquisition of Lattice Semiconductor by Canyon Bridge Capital Partners. Donald Trump didn't allow it to happen. Lattice was like #3 in FPGAs and even then the sale did not happen because of US govt.

Click to expand...

That's exactly what I'm saying, US abused this national security issue first in the context of mergers and acquisitions in semiconductor companies. China should do the same back and block this merger, the loss for China if AMD/Xilinx merger failed is much small than that of the US.

Look at the consequences if the merger was not approved by China. If an independent Xilinx downsizes, some talent will flow back to China. If Xilinx wants to survive on its own, it has to find ways to get around US chip bans and continue selling to China. If Xilinx is forced to get subsidies just to survive then the industry will become anti-competitive, which is also not in the best interest of US.

And what would be the point in that? I fail to see how spoiling the merger would push forward China's case anyway.