The concept of AI PC is gradually gaining popularity. CounterPoint data shows that about 45% of notebooks shipped in the first quarter of this year are AI PCs. This has successfully led the global PC market to stop the decline for eight consecutive quarters and finally rebounded and grew by about 3%.
The industry's definition of AI PC has gradually become clearer. The three major engines of CPU, GPU, and NPU are considered necessary, and each has its own strengths.
Advertisement
Among them, the CPU is suitable for general-purpose loads and is the most flexible. The emerging NPU is suitable for continuous low loads to greatly save power consumption. However, the scope of application is still relatively limited, and the computing power cannot be too high.GPU has always assumed core responsibilities with its unparalleled powerful computing power.
Naturally, NVIDIA has the most say in this regard. On the one hand, professional-grade GPU accelerators have almost monopolized the entire industry. On the other hand, consumer-grade GeForce RTX 40 series GPUs are revolutionizing experience in all walks of life.
In 2018, NVIDIA launched the new GeForce RTX 20 series GPU and RTX technology, which for the first time provided comprehensive acceleration for AI through a dedicated Tensor Core. DLSS super-resolution technology is the most advanced and intuitive embodiment of AI in games.
Advertisement
Even the entry-level GeForce RTX 4050 laptop GPU delivers up to 194 TOPS of computing powerGeForce RTX 4090 D can provide up to 1177 TOPS of computing powercovering all kinds of desktops, notebooks, and workstations. However, the current lightweight AI PC without independent GPU has a computing power of only about 10-45 TOPS, and the next generation can break through 100 TOPS.
On top of hardware, NVIDIA RTX also provides an extremely rich set of AI-related development kits and application software acceleration. In particular, TensorRT can greatly accelerate the very popular AIGC models, including the most popular Stable Diffusion 1.5, SDXL 1.0, etc., and has also received new Support for UL Procyon AI image generation benchmark.
Currently, RTX AI PC has more than 100 million users and more than 500 applications and games, forming a rich ecosystem.
Recently, NVIDIA held a special RTX AI media tasting event, announcing a number of AI-related content updates, displaying a number of AIGC artistic creations, and demonstrating the latest technology DEMO.
[NVIDIA RTX AI: Everywhere]
With the joint efforts of the entire industry, AI technology, especially AIGC, is becoming extremely popular and penetrates into all walks of life. In summary, it can be divided intoTen major fields including painting, graphic design, 3D creation, video editing, video experience, conference/live broadcast, document assistant, programming development, games, and game development.
In every field and even every scene, NVIDIA RTX GPU is providing various innovative acceleration technologies.
Stable Diffusion Drawing Painting, RTX GPU is the fastest acceleration method on PC. RTX 4090 D can generate 8 images per second by default. After TensorRT acceleration and further performance optimization, it can generate more than 120 images per second.
RTX GPU is also the most productive acceleration method, supporting all SD WebUIs on the market (Automatic1111, ComfyUI, SD-Forge, Qiuye, etc.) and the most advanced SD workflows (LoRA, ControlNet, fine-tuning, AnimateDiff, IPAdapter, InstantID, real-time painting, etc.).
graphic designfully supports the mainstream design software on the market, and can provide acceleration for various image editing functions.
Such as smart portrait, object selection, detail enhancement, image super-resolution, picture restoration, color restoration, RAW enhancement, sky selection, eye enhancement, etc.
come3D designRTX 40 series GPUs can provide AI acceleration for various offline renderers and real-time renderers. The recently popular D5, the famous UE, and Unity can all benefit from it, and DLSS can be used among them.
In terms of video editing, it also supports mainstream software tools and has quite rich functions, such as object masking, automatic scene detection, automatic picture reconstruction, video super-resolution, sharpening mask, color correction, speech recognition, human body tracking, and style conversion. , video frame filling, etc.
Video experienceIn terms of performance, the most noteworthy ones are RTX VSR super resolution and Video HDR, which can greatly enhance the image quality experience of online videos.Supports playing any video in Chrome and Edge browsers, and supports domestic mainstream video websitesand the driver has its own functions, so there is no need to download any additional software or plug-ins.
Whether it is the latest RTX 40 series or the early RTX 30/20 series, you can enjoy this benefit.
Conferences and Live BroadcastsIn terms of performance, NVIDIA Broadcast provides a perfect solution, which can bring the best streaming quality optimization through eye contact, face tracking, background blur, auto-focus, video noise reduction, vignetting, echo cancellation, etc.
It supports almost all video conferencing software and live broadcast software on the market, includingTencent Conference, DingTalk, FeishuTeams, OBS, etc., all RTX 40/30/20 series can be used.
human-computer interactionOn the other hand, AI dialogue and document organization are typical applications, and NVIDIA has also created a ChatRTX technology demo. , used for local enhanced retrieval generation.
It can use local RTX 40/30 series GPU (8GB and above memory) computing power to accelerate, perform large language model inference, intelligently offline retrieve local privacy documents (txt/docx/pdf, etc.) and generate results, especially supporting Chinese Dialog models and document retrieval, including Zhipu ChatGLM3, etc. Developers can also add other large language models that support TensorRT-LLM acceleration.
More importantly, ChatRTX is free and open source.
developIn terms of aspects, TensorRT can be said to be the fastest deep learning inference acceleration framework currently on the market, especially it supports leading Chinese open source models and can run on RTX 40/30/20 GPUs.
In terms of inference performance, compared to LlaMA.cpp, its generation speed can be up to 4 times faster. The larger the Batch Size, the more obvious the advantage.
On Stable Diffusion's raw graphs, compared to the general-purpose PyTorch, the speed can easily be doubled, and it can easily crush Apple's CoreML framework.
The most typical one in game applications is the combination of real-time ray tracing and DLSS. You can choose different quality and speed according to your needs. I won’t go into details about this.
A typical example in games is NVIDIA RTX Remix, which can revitalize old DX8 and DX9 games, automatically capture game scenes and materials, enhance them through AI, and add real-time ray tracing and DLSS 3/3.5. Portal RTX version is the best example.
In terms of hardware, it supports RTX 40/30 GPU.
In fact, NVIDIA has provided end-to-end development platform support for AI development in all fields. Development, training, integration, and deployment can all be accelerated by RTX GPUs.
Omniverse, Nemo, DLSS, ACE, Optix, VSR, Maxine, TensorRT, CUDNN… I am afraid this is the only one with such rich development tools.
Editor in charge: Shang Shangwen Q