NVIDIA's biggest competitor has emerged.

Author | Zhou Zhiyu Editor | Zhang Xiaoling's price rises 30% a week, and the price of a single piece exceeds 100,000. The heat of artificial intelligence and the shortage of chips have made the Nvidia A800 a hot hard currency in the market. Taking advantage of this east wind, Avida CEO Huang Renxun became one of the most high-profile entrepreneurs side by side with Musk. Nvidia's market capitalization also soared to $1.17 trillion, making it the hottest tech company on the market. However, on the one hand is the scarcity of high-end AI chips, on the other hand is the "100-model war", the downstream demand for computing power is high. In the face of a market that is just beginning to take off, competitors, large and small, are struggling to catch up. They hope to break the current pattern of "AI chips = Nvidia" and get a share of this trillion-level market. Just on July 11, Intel released Gaudi 2, an AI processor specially designed for the Chinese market, in Beijing. It is a benchmark for Avida 100 series and is specially built for training large language models. The launch of Gaudi 2 means that there is another giant in the AI chip market. Before Intel, AMD also launched its own AI products. Nvidia, Intel and AMD, the three chip giants that have been fighting hard since the PC era, are also facing each other in the AI era. There will not be a single AI chip market. With the entry of a giant, a brand-new competition has begun. ## Confrontation With the launch of Gaudi 2, Intel launched a frontal attack on Nvidia. The shrinking PC market and the weakening data center business put pressure on Intel's performance. Intel's market share, which was originally the "first brother" of the server chip market, was eroded by competitors such as AMD. The wave of artificial intelligence brings the demand for computing power, which makes Intel see a new power point. Designed by Habana Labs, an AI startup acquired by Intel for $2 billion in 2019, Gaudi 2 was built from the beginning of its launch to improve the efficiency of deep learning training in the cloud and data centers. At the press conference, Sandra Rivera, Intel's executive vice president and general manager of the data center and artificial intelligence division, spent a lot of time introducing the performance of Gaudi 2, comparing Avida's high-end GPU A100 and H100. From the demonstration data, such as Bert model pre-training, the performance of Gaudi 2 is 1.7 times that of Nvidia A100. As for the more advanced Nvidia H100, Eitan Medina, chief operating officer of Habana Labs, said that Gaudi 2 is currently the "only alternative" that can replace Nvidia H100 for LLM training ". In the MLPerf 3.0 benchmark, only Gaudi 2 and H100 were able to train GPT-3. From the current data, based on GPT-3 model training, there is still a certain gap between Gaudi 2 and H100. The performance of a single H100 is 3.6 times that of Gaudi 2. However, Eitan Medina said that with Intel's release of software support and new features for FP8 in September, Gaudi2 is expected to surpass H100. Cost performance is a core advantage of Gaudi 2 against Nvidia 100 series. Eitan Medina told Huawei that the performance per watt of Gaudi2 is about twice that of Avida A100 when running ResNet-50, and the performance per watt of the 176 billion parameter BLOOMZ model is about 1.6 times that of A100. **In other words, while providing good performance, Gaudi 2 is significantly better than Nvidia A100 in energy consumption, and Gaudi 2 can also challenge H100 in terms of cost performance. Intel has also become the most competitive opponent among the current challengers of NVIDIA. **Although similar to the A100, Gaudi 2 is different from the international version in order to comply with the relevant regulations of the US Bureau of Industry and Security. However, Eitan Medina said that the overall performance of the Chinese version of Gaudi 2 is not much different from the international version. The 5nm Gaudi 3, which is planned to be launched next year, will also be available to Chinese customers in compliance. At present, Intel has cooperated with domestic server manufacturers such as wave information, Xinhua three and super fusion, as well as Baidu intelligent cloud and other companies. Liu Jun, vice president of wave information and general manager of wave AI & HPC product line, also said that Intel has jointly released a new generation of AI server NF5698G7, supporting 8 Gaudi2. In addition, Sandra Rivera revealed that Intel will integrate Gaudi's AI chip and GPU product lines by 2025, when a more complete next-generation GPU product will be introduced. Through a wide range of product lines, to meet a variety of different needs. ## The race to Intel is not the first chip giant to charge Nvidia. In June last year, AMD also launched the CPU + GPU architecture Instinct MI300 to enter the AI training market. Then in June this year, AMD offered MI300X with up to 192 GB HBM memory to further optimize for large model training. AMD's data center hardware director Forrest Norrod said that the AI boom led by ChatGPT surprised AMD. The industry is still eager to Nvidia has a competitor, in addition to Nvidia's chip, there is an alternative option. And so it is. The growing demand for large model training, and limited capacity, makes Nvidia full of "happy troubles". Nvidia revealed that its orders have been scheduled to 2024, and H100 will be sold out before the first quarter of next year. At the World Artificial Intelligence Conference (WAIC) not long ago, Wang Yu, director of the Department of Electronic Engineering of Tsinghua University, also emphasized the current shortage of computing resources. He said that the high cost of deployment, the large gap in model computing power and the need for expansion and construction of domestic chip ecology are the three major challenges for the current large model landing. The continued growth in demand for computing power, as well as the desire of large model players to reduce the gap with the computing power support of manufacturers such as OpenAI by using better products, have made the demand for computing power high in the market. Players in the semiconductor market are also facing new opportunities. Sandra Rivera said that in the first quarter, demand for chips brought about by artificial intelligence at least doubled demand for Intel's products, including Gaudi. In addition, the fourth generation Xeon processors also have a good market response when AI use cases and market demand are exploding. According to Morgan Stanley's prediction, the annual sales volume of the global AI computing semiconductor market this year, including NVIDIA and AMD's GPU, AI computing special chips and the outsourced production of these chips, will be about US $43 billion. Within four years, the global AI computing semiconductor market sales will reach $125 billion. This is a bright spot in the global semiconductor consumer electronics sales slowdown, revenue decline is expected to be a bright spot. This trillion yuan market has naturally attracted large and small players to flock in one after another. They believe there is more than one Nvidia in the market. Nvidia does have its own barriers. GPUs have obvious advantages over CPUs in parallel computing power, memory bandwidth and other performance and floating-point computing speed. Nvidia also takes this advantage to take the lead in model training and reasoning, leading players such as AMD and Intel. Nvidia Unified Computing Design Architecture CUDA also relies on a closed ecosystem to form its own software ecosystem, binding millions of developers. Sandra Rivera also admitted that many people are using CUDA when conducting artificial intelligence and AI operations. However, in her observation, many developers of large models will not do such low-level development, but will innovate in PyTorch and TensorFlow. This is also an opportunity for players such as Intel. Sandra Rivera believes that software development or developer ecology has always been Intel's strong point. In the data center field, in addition to CUDA, it is Intel's X86 software ecology. Intel, on the other hand, wants to provide more competitive options for customers who want to move away from closed ecosystems and seek efficiency and scale. In addition to NVIDIA, AMD and Intel three chip giants, there are still big players in the market, ready to go. Mark zuckerberg's Meta Platforms announced on July 18 that it would join hands with Qualcomm to use Qualcomm chips to run Meta's new large language model Llama 2 on mobile phones and personal computers. Musk also said at the xAI conference that he was ready to develop his own AI chip. In addition to the fiery "100-model war", the battle on the hardware side has also begun. The traditional chip giants and the emerging new players are fighting fiercely on AI. Now the wave is beginning, and a battle for kings in a new field has begun.