Creating the future of general-purpose computing? Tianshu Intelligent Chip "DeepSpark" debuts

In recent years, the arrival of 5G and AI has promoted the rapid development of the digital economy. All walks of life require stronger computing power. The increasingly diverse application scenarios require cloud computing resources to cope with the needs of various complex scenarios and can provide Rich and flexible computing support, in this context, general-purpose GPU has gradually become the current mainstream trend.

According to data from market research organization Verified Market Research, China's general GPU chip market reached 8.6 billion yuan in 2019. It is expected that by 2025, the market size is expected to reach 45.8 billion yuan, with more than five times the growth space. Faced with such strong market demand, how to distinguish true general-purpose GPUs has become a difficult problem for people to face. On August 30, the Tianshu Zhixin Top 100 Application Open Platform Conference was successfully held, officially launching the DeepSpark open platform.

"Currently, the versatility of chips has become a very important thing." Lu Jianping, CTO of Tianshu Zhixin, mentioned in his opening speech. He pointed out that the versatility of chips must have two major aspects. One is the chip itself. The architecture must be universal, and secondly, it needs customer support. Compared with the architecture of the chip itself, customer support may be more important. Take Tianshu Zhixin's Tiangai 100 with a universal architecture as an example. So far, 300 customers have jointly defined what universality is. Regarding the "top 100 applications", Lu Jianping explained that currently there are more than 300 Tianshu Smart Core customers, and the application adaptation scenarios have reached 200 scenarios. The top 100 applications are just a starting point for Tianshu Smart Core, not the end.

Lu Jianping, CTO, Tianshu Zhixin Chief Technology Officer

What is chip versatility? Lu Jianping said in an interview with the media that the current versatility can certainly be applied to more scenarios, but in fact time is also included in the cost. If the chip takes a long time to adapt to the hardware, it is not really universal. Versatility is a practical concept, at least it can be adapted to the customer's hardware with reasonable capabilities and resources.

Academician Ni Guangnan of the Chinese Academy of Engineering pointed out in his speech that in the era of digital economy, computing is productivity, which has become a global consensus. Facing the increasingly strong computing demand, the current traditional computing chips based on chips can no longer afford the geometric increase in computing power, and chips represented by general-purpose GPUs are rapidly emerging.

Academician Ni said that Tianshu Zhixin's top 100 open platforms have opened up their technical capabilities to the industry, which will not only set a benchmark for other computing chip companies, but will also provide strong support for the independent ecological construction of my country's computing products. In this regard, Academician Ni put forward two suggestions for the Tianshu DeepSpark open platform:

First, it is hoped that with the help of such an open platform, Tianshu Zhixin can form a close collaborative relationship with upstream and downstream enterprises to jointly develop adaptation systems and solutions, achieve application in a wider range and more scenarios, and accelerate the construction of an intelligent computing industry. ecosystem.

Secondly, I hope that Tianshu Zhixin will continue to open up its platform and open source the platform in the future so that more companies can participate and contribute. We should also learn from Tianshu Zhixin and open up and open source our capabilities to the industry to jointly help the computing industry. High-quality development.

It is understood that Tianshu Zhixin released the first domestic 7nm general-purpose GPU for cloud training, Tiangai 100, in March last year, achieving a breakthrough in domestic general-purpose GPU chips from 0 to 1. Over the past year or so, Tiangai 100 orders have The total has exceeded 200 million yuan; in May this year, Tianshu Intelligent Core successfully lit up the "Smart Armor 100"; now, Tianshu Intelligent Core has successfully held the top 100 application open platform launch conference. In the view of Academician Ni, the top 100 models reflect the product's Universal.

At the meeting, Zou Yi, vice president of Tianshu Zhixin, introduced “DeepSpark” in detail. Zou Yi mentioned that as the first domestic company to achieve mass production of general-purpose GPUs, Tianshu Zhixin realized the strong domestic demand for general-purpose computing power during the deployment of general-purpose computing power: on the one hand, the sales of mainstream products in the market are constantly rising. , prices have hit new highs one after another; on the other hand, Tianshu Zhixin has discovered more than once that everyone is confused about general computing power.

Zou Yi, Vice President of Tianshu Zhixin

How to use a faster way to verify the versatility of computing power based on common algorithms? Tianshu Zhixin provides a solution based on rich applications - DeepSpark Top 100 Application Open Platform.

The platform will be launched on GitHub and Gitee on September 30, when Mindstore and PP Flying Paddle will join. Zou Yi emphasized that Tianshu Zhixin and DeepSpark partners will fully invest in the construction of the DeepSpark platform from now on. In the next year, we hope that users can see that its applications can be used to create two, two, three, or endless, empowered by computing power. A better future.

It is reported that the DeepSpark platform will be operated and maintained in an open source community manner. Tianshu Zhixin will share the implementation results jointly built with partners and community developers, and provide community developers with an application development environment within reach, empowering more During the development process, many industry partners are exposed to more applications through algorithms, thereby changing daily life. At the same time, the DeepSpark platform will also gather the power of community developers to help hardware and software partners in the industry iterate products through co-construction, and improve the entire general computing ecosystem more quickly.

According to Zou Yi, the DeepSpark platform will mainly provide three different parts in the open source community:

The first part is application algorithms. In application algorithms, in addition to the algorithm models used in daily AI, the DeepSpark platform will also widely accommodate implementation algorithm models in various fields such as 3D reconstruction, general computing, and high-performance computing. In the first version, DeepSpark will provide more than 100 models and algorithms. Tianshu Zhixin has selected algorithms for the top 100 applications. The existing algorithms have covered various fields from AI images, vision, and natural speech, general computing outside the field, and super-resolution.

Taking classification algorithms as an example, DeepSpark covers the algorithm evolution accumulated in this field over the past decade and supports the latest CVPR algorithms in emerging fields. Every year, it supports the latest algorithm models based on the latest CVPR trends and the above rankings. In addition, DeepSpark will also support the latest large domestic speech analysis models in the language field, including new models such as CPM and CRO.

Simply put, DeepSpark will propose a general-purpose GPU or general-purpose computing power algorithm set to the industry for the first time.

The second part is the framework area. The framework is equivalent to an operating system for application development. DeepSpark is directly compatible with mainstream international frameworks in the industry, as well as leading domestic frameworks. At the same time, DeepSpark also provides developers with a series of complete toolboxes. Whether the development comes from voice, video or general computing fields, you can find toolboxes commonly used in daily work, making development more convenient.

Zou Yi pointed out that the DeepSpark platform mainly hopes to empower all application developers to directly upgrade and debug under existing development, and to complete rapid cross-platform comparisons. In this way, DeepSpark can solve the pain points of difficulty in cross-platform migration and lack of tools for application development.

In the third part, DeepSpark will provide a one-stop, multi-dimensional evaluation system. Not only is it cross-OS, the DeepSpark platform also helps infrastructure construction companies choose their own computing power platform, and provides a series of evaluation references for the industry based on the applications that each industry needs to implement.

"Ultimately, the DeepSpark platform, as an open source community, is a community that is completely independent of hardware. In other words, everyone can use a variety of hardware to build their own evaluation system to understand the degree of application and actual versatility." Zou Yi emphasized.

It is worth noting that Lu Jianping also pointed out in an interview that DeepSpark has nothing to do with the brand. In theory, non-GPU people can also use DeepSpark to test hardware. Users of any AI chip manufacturer can use DeepSpark to evaluate the entire hardware. Although it is currently difficult for general AI hardware to run through DeepSpark, in the future non-GPU and GPU manufacturers can use the DeepSpark platform to test and develop hardware, making the entire hardware more versatile. "DeepSpark is a mechanism to help everyone more centralize their computing power," Lu Jianping said. In addition, Lu Jianping also revealed that open source is a trend in the world. What is currently lacking in China is versatility. In the future, based on business needs, strategic needs, and help for the company’s future development, in addition to the open source community, Tianshu Zhixin will also make Some options.

In addition, at the meeting, Ding Cheng, general manager of Shengsi's business, shared Shengsi's MindSpore open source strategy and prospects for cooperation with Tianshu Smart Chip. Ding Cheng pointed out that as a late-developing framework, Shengsi will build its own unique competitiveness and unique advantages from the following three aspects: First, Shengsi's business has full-scenario deployment capabilities at Huawei's own software and hardware equipment vendors. It can be easily deployed to various equipment vendors; secondly, MindSpore continues to extend and support a series of parallel optimization technologies such as data parallelism and model parallelism, which can maximize the computing power advantages of the incubation cluster; the MindSpore team is also committed to building With its own native academic ecosystem, the entire MindSpore ecosystem has begun to take shape.

Baidu Flying Paddle team leader Mr. Zhao also shared the latest progress of Flying Paddle. Flying Paddle open sourced Paddle in 2016. Currently, more than 30 manufacturers at home and abroad are working on Paddle adaptation, including Intel, Cambrian, etc., as of now At present, Paddle has completed the adaptation work with six domestic CPUs of four architectures, including Intel and AMD.

write at the end

With the support of emerging fields such as AI and the Metaverse, people's demand for computing power has become increasingly strong. Tianshu Zhixin hopes that the DeepSpark Top 100 Application Platform can help everyone better identify more effective computing power in the computing business. I believe that in the near future In the future, Tianshu Intelligent Core will rely on these computing power to create a new era of general computing applications.

*Disclaimer: This article is original by the author. The content of the article is the personal opinion of the author. The reprinting by Semiconductor Industry Watch is only to convey a different point of view. It does not mean that Semiconductor Industry Watch agrees or supports the view. If you have any objections, please contact Semiconductor Industry Watch.

Today is the 3148th content shared by "Semiconductor Industry Observation" with you. Welcome to pay attention.