Eastern US Time2month21Today, NVIDIA released2024The results of the fourth fiscal quarter of the fiscal year show that both total revenue and data center revenue within the quarter reached record highs.2024FY revenue has also been achieved609Record $10.2 billion performance. The financial report shows that NVIDIA's revenue for the fourth fiscal quarter221billions, month-on-month growth22%A sharp year-on-year increase265%Net profit123billions of dollars, a sharp year-on-year increase765%Recently, there have been reports indicating that, amidst existing strengthsGPUBeyond the field, NVIDIA is preparing to enterASICDomain competition with BOTT.GPUenterASICIs there a significant threshold in the field? What does NVIDIA's new move mean? Amid the strong demand for large models,GPUbesidesTPU,LPUDoes the rise of these routes mean new possibilities for high-speed computing chips? What are the advantages and disadvantages of these routes?
Frost & SullivanFrost & SullivanMr. Lu Jing, Partner and Managing Director of Frost & Sullivan's Greater China Region, was interviewed21Century Economic Report interviews were conducted to jointly discuss the aforementioned topics.
21Century Economic Report
*Click on the end of the article to read the original text and view the complete report.
Q: Recently, there have been reports indicating that, in addition to existing strengthsGPUBeyond the field, NVIDIA is preparing to enterASICDomain and Broadcom competition. What's your take on it fromGPUenterASICAre there significant barriers in the field? What does NVIDIA's new move mean?
For the implementation of artificial intelligence, algorithms are at the core, and computation and data are the foundation. In terms of computing power, the main method currently used isGPUParallel computing neural networks, at the same time,ASICIt has also gradually become a more popular choice in recent years.ASIC (Application Specific Integrated CircuitsSpecial-purpose integrated circuit)Integrated circuits refer to integrated circuits that are designed and manufactured according to specific user requirements or the needs of specific electronic systems.ASICAs a product of the close integration of integrated circuit technology with the overall machine or system technology of specific users, compared to general-purpose integrated circuits, it has advantages such as smaller size, lower power consumption, improved reliability, enhanced performance, increased confidentiality, and reduced cost. UnlikeGPUandFPGAflexibility, customizedASICOnce manufactured, it cannot be changed, which makes the entry threshold high due to the high initial cost and long development cycle.
NVIDIA, as a globally dominant AI chip designer and supplier, is preparing to enterASICThe domain aims to capture a portion of the explosive market for custom AI chips and protect itself from an increasing number of companies seeking to replace its products. NVIDIA toASICThe entry into this field is likely to erode Bortronics and MaimianASICMarket share in the field.
Q: NVIDIA has also recently launchedChat With RTX, which seems to be mainly targeted atAI PCA product on the app market. It seems that NVIDIA has been relatively quiet in making moves towards the application side for some time. What do you think of NVIDIA's new move this time?
Chat With RTXIt's a personalizedAIChatbot, which can bePCRun locally to process videos, analyze files, and provide users with relevant answers based on their own data.
sinceChatGPTAfter the fire,OpenAIFounder Sam Altman is planning to raise a significant amount of capitalAIThe chip industry aims to solveAIThe supply and demand issue of chips. NVIDIA's new move this time is a concrete response to its role asAIPioneers in the chip industry, with expertise in the field of large models, will adopt a strategic layout combining software and hardware.AImarket
Q: A fresh face latelyGroqEmerging from nowhere, adoptingLPUThe way andGPUCompetes in computing models, reportedly leveragingSRAMrather thanHBMMakes computing very fast. Based on the information released so far, what do you think is the future of large models under strong demand?GPUbesidesTPU,LPUDoes the rise of emerging routes mean new possibilities for high-speed computing chips? What are the advantages and disadvantages among these routes?
GroqofLPU（Language Processing Unit) can bring higher computing speed while consuming less power. Setting aside the accuracy factor of the model,LPUThe computing speed in specific domains is currently widelyGPUCompared to being able to be fast10More than times.LPU,TPUThe rise of high-speed computing chips meansAIThe implementation of large models will be further accelerated.
TPU,LPUandGPUThe biggest difference is their architecture and application domain.GPUSuitable for general computing tasks, it has parallel computing capabilities and a large-scale program cache, such as in graphics rendering, physical simulation, and numerical computation.LPUandTPUFocusing on the acceleration of deep learning and machine learning applications, it offers lower power consumption in specific domains. Users can choose the appropriate processor based on different computing task requirements to improve computing performance and efficiency. Considering the comprehensive cost factors of actual operation, it offers stronger general performance.GPUIt remains the preferred choice for training large models.
Q: More and more large model vendors (such as Google, Microsoft,OpenAI(Optional) selection/Consider preparing for self-developed solutionsAIChip, plus the old rivalAMDThe tiger is eyeing covetously. Do you think this could accelerate a change in NVIDIA's dominant position (for example, towards a one-supremacy, multiple-strongholds scenario)? If not, what are the main constraints?
The explosion of large models has propelled NVIDIA to become the most discussed topic todayAIChip companies have also seen new highs in revenue and net profit disclosed in their annual reports. However, asAIThe craze continues to heat up, and more and more large models are starting toAIThe chip industry is making efforts to break free from NVIDIA's supply constraints as soon as possible. Large model vendors want toAIThe volume of operations has increased significantly, requiring the support of proprietary chips to achieve optimal R&D efficiency. Therefore, more and more large model manufacturers are choosing to develop their own chips.AIchip. On the other hand,AMDAlso23year12Launched in JanuaryAIChip, embarking on a head-on competition with NVIDIA.
However, these manufacturers cannot shake NVIDIA's dominant position in the short term. Firstly, NVIDIA has an undisputed first-mover advantage and rich industry experience; its chips have been widely used in major model manufacturers. Secondly, NVIDIA has built a comprehensiveCUDAecology,CUDAIt is based on the launch of NVIDIAGPUThe parallel computing platform and programming model can be used to accelerate large-scale data parallel computing, enablingGPUCan be used in a wider range of scientific computing and engineering computing fields. After years of development, NVIDIA'sCUDAalready has400Ten thousand developers have essentially formed an ecological barrier that is monopolistic. The software ecosystem is precisely the product competition element that downstream customers attach the most importance to. If one rashly changes the ecosystem, it means that manufacturers will incur increased learning costs, trial-and-error costs, and debugging costs. Moreover, NVIDIA has further consolidated its moat through investment.2023Since the beginning of this year, NVIDIA has invested in more than twenty companies, ranging from large-scale new artificial intelligence platforms worth billions to small startups applying AI to industries such as healthcare or energy, all of which have formed close connections with NVIDIA.
For large model companies' self-developed chips, the most critical bottleneck lies in the inability to resolve supply chain issues in the short term. The production capacity of foundries and wafer fabs is already saturated. Considering comprehensive costs, self-developmentAIIt is not necessarily more advantageous to purchase chips externally.AMDofAIChip, NVIDIA'sAIThe chip is more suited to potential customers with abundant funds and high performance requirements. In the short term, NVIDIACUDAThe ecosystem remains robust, and most users who need training chips still choose NVIDIA.
* This interview was published in 21Finance and Economics The reporter is Luo Yiqi, and the original title is: NVIDIA's All-Out Offensive: Record-breaking Performance, Hardware and Software Collaborations |GAIEvolutionary Theory ⑤
