In early September, as the United States restricted exports of high-performance data center GPUs to China, the trend toward domestic acceleration computing became increasingly clear. Domestic chip companies are seizing the opportunity to grow, driving heat in the AI chip and GPU markets. In the capital market, related Chinese concept stocks rose in response.

"In the AI chip field, the timing, location and people are all favorable, including national policies, talent and experience - this is a great opportunity." Recently, Zhao Lidong, founder and CEO of Enflame Technology, said in an interview with Jiemian News that the key to seizing this opportunity is to break the ecosystem monopoly and achieve architectural innovation.

Enflame Technology was registered and established in March 2018. After the seed round, it publicly announced four rounds of financing, all of which were participated in by Tencent, with cumulative financing of nearly RMB 4 billion. It achieved productization of AI chips early on, with a product portfolio covering cloud-based AI training and inference chips and their accelerator cards, as well as AI training OAM modules. It is one of the earliest domestic AI chip startups to have both cloud training and cloud inference solutions. At the World Artificial Intelligence Conference (WAIC) in early September, Enflame Technology released the "YunSui AI Computer" AI accelerator cluster product, entering the cluster computing field.

Enflame's founding team comes from AMD. Zhao Lidong previously served as Vice President of Unigroup, and earlier worked at AMD for 7 years, holding positions including Senior Director of the Computing Business Unit. Another founder and COO, Zhang Yalin, worked at AMD's Shanghai R&D center, serving as R&D lead for multiple chip projects.

The complete team structure is an important reason why investors favor Enflame. Besides the founders, a significant portion of Enflame's senior talent also comes from AMD. "AMD once derived one-third of its global revenue from China, and China has a complete structured R&D system, but lacks product definition authority," Zhao Lidong explained. Product definition means coordinating various departments to determine chip products, define specifications, conduct competitive analysis, and clarify market prospects. This important work has always been undertaken by AMD's headquarters in Austin, Texas.

The AMD experience and the founding team's chip development expertise became the opportunity and foundation for Enflame's establishment. In terms of products and team size, over nearly four years, Enflame grew from fewer than a dozen people to 800, its products evolved from a single training chip to end-to-end solutions, and it eventually ventured into more complex cluster computing.

Zhao Lidong said the opportunity Enflame sees is: under the growing demand for AI computing, traditional CPUs (central processing units) can no longer meet computing power needs, and GPUs (graphics processing units) have leapt to become the best choice thanks to their parallel computing capabilities. NVIDIA's complete software ecosystem of math libraries, operators and function libraries built for GPUs has made GPUs the undisputed protagonist of AI computing.

AI computing is roughly divided into two levels. First, training the model, which can take days; then the trained model responds to actual requests, such as recognizing an image and making inferences. Architecturally, about 60% of NVIDIA's traditional GPU area is used for general computing and 30% for graphics rendering; GPGPUs (general-purpose computing GPUs), mainly used for accelerated computing, generally remove the graphics rendering portion. Over the past decade or more, NVIDIA captured the wave of AI technology rise and built a software ecosystem through the CUDA platform, making its GPUs the standard for processing AI applications in data centers worldwide, holding over 90% market share. In the cloud domain, NVIDIA's GPUs currently hold significant share in the training market, while some inference tasks are handled by Intel CPUs.

The industry generally believes that training chips are harder to develop and commercialize because during AI training, neural network models are not fixed, creating a high demand for chip generality, making NVIDIA GPUs the long-standing first choice. Inference is simpler and grows faster, so training chips impose a higher test on chip companies' design capabilities. Looking at the global AI chip market, inference-first-then-training is the mainstream path; Intel's acquired AI chip company Habana and many domestic AI startups follow this route as well.

From this perspective, Enflame Technology's route of launching training chips first and inference chips later is quite distinctive.

In December 2019, Enflame Technology released its first cloud training chip "Suisi", deriving two computing cards for training and inference from this chip. "Not just making big chips, but making AI training chips, which are far more complex than inference. Training chips can do inference, but inference chips cannot do training," Zhao Lidong said. Enflame subsequently launched the second-generation training chip Suisi 2.0 and inference chip Suisi 2.5, totaling three chips.

In fact, in the AI computing field, NVIDIA's dominant position is expected to be changed by the impact of many cloud giants and startups.

Especially in China, the powerful NVIDIA is not without "weaknesses": the vast majority of AI vendors rely on NVIDIA's software and hardware ecosystem, lacking bargaining power; on the other hand, developing AI features based on NVIDIA's non-open-source software makes deeper development difficult. In terms of customer support and service teams, NVIDIA finds it hard to treat Chinese customers like BAT equally with American companies such as Google and Amazon.

Therefore, it is necessary to build a self-owned chip ecosystem supply chain. According to Zhao Lidong, Enflame chose to build its own development toolchain adapted to its own hardware. However, most domestic chip makers specializing in AI currently rely on or are compatible with NVIDIA's CUDA software ecosystem, thereby avoiding NVIDIA's software ecosystem barriers at the startup stage. Zhao Lidong believes it is difficult to evaluate this approach from an integrated software-hardware perspective; surpassing NVIDIA GPUs while remaining CUDA-compatible may involve intellectual property issues.

Enflame Technology is also trying to advance the integration of its hardware system and software, collaborating seamlessly with machine learning algorithms, such as adapting algorithms collected in machine learning repositories like Google's TensorFlow, Facebook's PyTorch and Baidu's "PaddlePaddle", to complete its ecosystem. The so-called AI chip ecosystem refers to software layers including drivers, compilers and developer toolkits that simplify the work of AI developers.

It is worth noting that startup chip companies like Enflame also face the question of how to deal with large internet companies.

Currently, Chinese internet giants are all positioning themselves in the AI chip field. BAT are also mainstream cloud service providers, purchasing large amounts of computing hardware every year. Alibaba's semiconductor division "T-Head" has already built an AI chip business, Baidu has the AI chip company Kunlunxin, and Tencent Cloud has disclosed progress on three self-developed chips, saying one AI chip has been taped out for trial production. Among the three, both T-Head and Kunlunxin have explicitly stated they will expand external customers.

In the process of chip deployment, cost is also a consideration in corporate decision-making. Whether in terms of cost, development cycle or maintenance, internet companies' self-developed chips are hard to optimize; some develop chips only for specific scenarios with limited internal scale, and since scale determines profitability, the economics are hard to justify - without profitability, sustained iteration is impossible.

Zhao Lidong emphasized that in the face of competition from other chip companies, Enflame must achieve the ultimate in chip cost-performance and energy-efficiency ratios. According to him, Enflame's main goal this year is large-scale product deployment and expanding market share. The third-generation chip products will be taped out by the end of the year, and profitability is expected on the basis of second-generation product scaling.

However, with Enflame's chips iterating continuously, upfront costs are enormous. Only by further increasing revenue and reducing costs after scaling, and improving gross margins, can profitability be achieved. "The time (to achieve it) depends on the market environment. If market share remains the priority, it may be necessary to sacrifice gross margin and lower prices to compete, and it also depends on demand. Ultimately, a comprehensive judgment decision must be made after considering multiple factors," Zhao Lidong said.