AI is becoming a new variable in the PC industry.
Recently, AMD, Intel, and others have integrated NPU (Neural Processing Unit) units into their own processors to enhance AI processing capabilities. For example, Intel's recently released Core Ultra series processors, combined with NPUs, CPUs, and GPUs (graphics processors), can provide 34 TOPS of overall AI computing power. AMD released a new 8040 series AI PC processor earlier this month, with an overall nominal computing power of 39 TOPS and a built-in NPU computing power of 16 TOPS.
Both the industry and third-party institutions hold a relatively optimistic attitude towards the development trend of AI PC. According to IDC's forecast, the penetration rate of AI PCs will rapidly increase to 55% next year and reach 85% by 2027. The proportion of AI terminals among all terminals will also increase from 41% this year to 79% in 2027. With the launch of a large number of AI PCs next year, the overall price of PCs will significantly increase. However, Counterpoint stated that AI PCs are likely to drive a new round of shipment rebound next year and dominate the PC market after 2026, with an expected penetration rate of over half by then.
But when processor manufacturers such as Intel and AMD attempt to integrate NPUs to roll up AI, they still need to deal with competition from another GPU giant, Nvidia.
As a proponent of Moore's Law, NVIDIA CEO Huang Renxun has always been pessimistic about the long-term prospects of CPUs. He believes that CPUs are facing performance bottlenecks, and the rule of increasing performance by 10 times every 5 years under the same power consumption in the past is no longer applicable. The industry needs a new computing method, which is accelerated computing driven by GPUs.
GPU was originally designed as a chip for electronic and computer games, and in order to process complex and realistic graphics, it was necessary to process 3D models, lighting and shadows, and other image details. Compared to CPUs that perform calculations in sequence, GPUs that can parallelize a large number of simple calculations are more suitable for this type of computing load.
Since the release of the RTX series GPU in 2018, Nvidia has emphasized that its products have AI technology capabilities such as ray tracing and DLSS oversampling: on the one hand, they can bring better image rendering capabilities in gaming, graphics rendering and other application scenarios; On the other hand, technologies such as DLSS reduce the computational power requirements brought by graphics processing, achieving a balance between image performance and power consumption.
At present, Nvidia's consumer grade GPU product architecture has been upgraded to Ada Lovelace, and it is reported that the overall performance of the graphics card has been improved by four times. Important performance improvements include the participation of AI processing units, such as DLSS3 technology, which can insert frames for games and improve screen smoothness. In terms of 3D design purposes, the Nvidia version of the "Metaverse" Omniverse relies on DLSS technology to output 4K high-definition images, etc.
In addition to graphics processing, the new market for GPUs has also expanded to fields such as scientific computing and data processing, among which Nvidia's parallel computing platform CUDA, released in 2006, played an important role. CUDA is a software tool that utilizes GPU resources, which includes a series of development tools. Before CUDA, developers had to use GPUs originally designed for image rendering to accelerate computing, which required complex debugging and compilation. After more than a decade of development, CUDA has become an AI infrastructure, which is a native AI development platform. Anyone who owns a Nvidia GPU computer with CUDA support can develop any AI related software.
For example, the AI voice changing software "VoiceMelody" on Steam can integrate and use phoneme information in speech, allowing for natural voice conversion regardless of speech quality or language, provided that the graphics card supports CUDA.
According to Huang Renxun, the download volume of CUDA software has reached 25 million times in the past year, while its historical download volume is 40 million times.
In the field of artistic creation, Chinese artist Zhao Enzhe told Interface News that due to technological barriers and industrial processes, creators used to have to compromise between conceptual design and final presentation. However, now with the help of Nvidia GeForce RTX graphics cards and AI creative tools such as Stable Diffusion and RUNWAY, conceptual designers can break through the limitations of creative categories, Try many possibilities of implementation in just a few seconds.
For generative AI, GPUs still have a first mover advantage, and currently mainstream large-scale model applications are built on GPU driven accelerated computing. In terms of software tools and support, Nvidia is also continuing to adapt and optimize applications such as Stable Diffusion.
Huang Renxun believes that the PC industry is facing a rebirth opportunity. In the next decade, new AI PCs will replace traditional PCs, with a market value of over trillions of dollars. But in order for major manufacturers to seize market share from Nvidia, they not only need to have stronger hardware computing power, but also need to put greater effort into the software ecosystem, while achieving the best in AI deployment, inference speed, stable operation, and compatibility.