Write at the beginning
In the past six months, in addition to the amazing effect of DLSS3 of RTX40 series graphics cards, AIGC's popularity has far exceeded players' imagination. In the past, no one thought that the powerful computing power of GPUs could be used for AI. Although compared with large-scale AI computing clusters, the computing power of game graphics cards is not worth mentioning, but we use it to implement some simple AI applications is quite good, of course, AI can also help you improve work efficiency, etc., these are very practical embodiment.
It is not difficult to realize AI applications, and now you only need a high-end computer to enjoy it. But since it is an experience, then how can we use an ordinary graphics card, we just received ZOTAC GeForce RTX4090PGFOC graphics card, as the strongest game graphics card on the surface, its AI ability must also be the focus of curiosity of major players. So how does this graphics card actually perform? Let's take a look at it together.
Introducing TensorCore
To experience AIGC, you must first understand the internals of the graphics card. At first, the game graphics card could not be used for AI training, and in the past it was more of a pure game graphics card. However, NVIDIA introduced TensorCore to the game graphics card, so the graphics card has greatly improved deep learning performance. This also makes AI another great use for gaming graphics cards.
The official name of TensorCore is Tensor Computing Core, which is used to increase the AI deep learning computing power of graphics cards. The AI application we are talking about this time is actually inseparable from this TensorCore. It can be said that since the addition of RTX20 series graphics cards, the graphics card has opened a new door, allowing players to not only play games, but also create more possibilities with AI.
The first generation of TensorCore
However, the first graphics card to implement TensorCore is not the RTX20 series of Turing architecture, but the familiar TitanV, as the only Volta architecture graphics card, it is the first to eat TensorCore. Before the release of RTX20 series graphics cards, many deep learning practitioners bought this graphics card for deep computing.
Speculatively, the TitanV graphics card is only equipped with 640 first-generation TensorCores, but it supports mixed-precision matrix multiplication under FP16 and FP32, providing deep learning performance of more than 100 trillion times per second (TFLOPS), more than 5 times that of the Pascal architecture. Compared to Pascal, peak teraFLOPS (TFLOPS) performance for training is improved by up to 12x, peak TFLOPS performance for inference is improved by up to 6x, and training and inference performance is improved by 3x.
Second generation TensorCore
The TitanV graphics card equipped with the first generation of TensorCore is actually not our conventional game card, really let this technology be decentralized, but also the Turing architecture RTX20 series graphics card, up to the flagship RTX2080Ti, down to the dessert-level RTX2060 have all introduced TensorCore.
It is precisely because of the introduction of TensorCore that RTX20 series graphics cards have certain deep learning capabilities, and Lao Huang can complete ray tracing and DLSS on RTX20 series graphics cards. The second generation of TensorCore is an improvement over the first generation, providing a range of precision for deep learning training and inference (from FP32 to FP16 to INT8 and INT4), providing up to 500 trillion tensor operations per second.
The third generation of TensorCore
In the RTX30 series graphics card, that is, the Ampere architecture graphics card, NVIDIA's TensorCore is upgraded to the third generation. Accelerate and simplify AI applications with new precision standards TensorFloat32 (TF32) and 64-bit floating point (FP64), which can speed up AI up to 20 times.
In the third generation of TensorCore, NVIDIA also introduced sparsity acceleration, which automatically identifies and eliminates less important DNN (deep neural network) weights while still maintaining good accuracy. First, the original dense matrix will be trained, the sparse matrix will be removed, and then the sparse matrix will be trained to achieve sparse optimization, thereby improving the performance of TensorCore.
So the end result is that the third-generation TensorCore processes sparse networks twice as fast as Turing, with a computing power of up to 238TensorTFLOPS, compared to Turing's 89TensorTFLOPS.
Fourth-generation TensorCore
The TensorCore on the RTX40 series graphics card has evolved to the fourth generation, the most important change is the addition of HopperFP8TransformerEngine, which can provide 1400TFLOPS tensor processing performance, it can be said that deep learning performance has been a huge leap, which also means that through it can achieve new technical ideas, later DLSS3 we will mention the credit of TensorCore again.
We talk about this FP8 low-precision floating-point feature, in fact, few people may pay attention to this hardware upgrade, but it has changed a lot in the AI field, and the biggest benefit after FP8 hardware acceleration is an emerging AI infrastructure in recent years, Transformer. In recent years, famous languages AI, such as BERT and GPT, have used this structure, of course, we are familiar with the field of AI mapping also use this structure.
The architecture of the Transformer model
This type of AI model is characterized by large and many parameters, and the FP8 data format can help them reduce the space occupation, cram more parameters into it, and calculate faster. We also mentioned above that the hardware structure responsible for FP8 acceleration in the RTX40 series graphics card of the Ada architecture is called HopperFP8TransformerEngine, and the Transformer is written in the name, which is enough to see its importance.
Of course, don't think that AI is only used in GPT, AI mapping and other fields, in fact, the DLSS 3.0 technology we talked about in the previous article is also a kind of AI application, because DLSS technologies such as multi-frame synthesis, complementing and super resolution are completed by relying on deep learning. Interested players can click [Hardware Chronicle] What is the use of DLSS technology, and can Vigor really produce miracles? Browse to see how Lao Huang pushed DLSS technology to the altar of today.
Introduction to the test platform
After reading the introduction of TensorCore, everyone is also curious about what kind of AI deep learning capabilities it can bring after four generations, right? Then we are not stingy, sacrifice the most TensorCore graphics card in the current RTX40 series graphics card - ZOTAC GeForce RTX4090PGFOC, to take you to see how such a top game graphics card performs wonderfully in AI.
First of all, the most eye-catching is its appearance, compared with the previous generation has a touch of elegance, the rounded shell design also has a lot more streamlined softness, injecting flow and rhythm into the graphics card, breaking the tradition of using sharp lines to outline the appearance of the graphics card.
The black and white color contrast design is both fashionable and avant-garde in terms of visual effects, but also brings a sense of calm and neutrality, quiet and balanced, but also conceals a hint of e-sports hardcore style, the design is quite advanced.
The flagship graphics card naturally has a flagship configuration, the backplane uses an all-metal structure, and the alloy reinforced bone structure components are also added, which is conducive to consolidating the graphics card, not easy to deform, and can also improve the heat dissipation efficiency, plus the end of the backplane is equipped with a hole design to effectively enhance the fan airflow.
The internal heat dissipation is also not vain, 3 fans 9 heat pipes luxury configuration, which is the treatment of high-end graphics cards. In addition, there are large areas of VC vapor chambers and dense heat dissipation fins, which are difficult for the core to heat up. It has to be said that only the flagship-level heat dissipation specifications are worthy of the positioning of the flagship PGF.
It is not enough to have heat dissipation, the power supply must also be full! This ZOTAC RTX4090PGFOC graphics card uses the SEP2.0 power supply system. The total number of power supply phases has come to an exaggerated 28 phases, of which 24 phases are core power supply, 4 phases are memory power supply, and the power supply integration is quite high, each phase uses solid capacitors as input and output, and fully enclosed inductors sit.
The external power connector is a new 12VHPWR power supply interface, and a 12VHPWR power cord can meet the power supply requirements. The 600W power supply capacity can fully feed this 530WTDP performance behemoth.
The video output interface is also a top-notch existence, after all, you have bought ZOTAC RTX4090PGFOC graphics card, how to say that you have to be equipped with a 4K display, real 4K to experience real gaming. The 3 DP1.4a and 1 HDMI 2.1 on ZOTAC on this card can well meet the video output, and support 4-screen display or the highest 8K@60Hz specifications are completely "overdrive".
Having said so much, the above luxurious configurations are for the ultimate performance release, and the source of performance is this AD102-301-A1 core, don't look at it is just a small chip, in fact, it benefits from the TSMC4N process, 76.3 billion transistors are stuffed inside, which is 2.7 times that of RTX3090Ti! Of course, it also has 384 bits wide, 144 SMs, 96MB L2 cache... Everything here tells you that with this core, you have the strongest gaming graphics card on the planet.
Without saying much, this is the actual measurement on the machine for everyone to see, as soon as the machine is on the machine, I witnessed the beautiful appearance of the ZOTAC RTX4090PGFOC. The ZOTAC ZOTAC Faith Logo light on the top lights up first, and the front light strip flashes synchronously, surrounding the three cooling fans, like a rhythmic note. The author is also deeply impressed by this card, with beauty and performance, worthy of the name of the card emperor.
Of course, the hardware used to test it can not be bad, the CPU uses the latest Intel Corei9-13900KS, Asus' Z790Hero motherboard with 2 16GB Kingston DDR5-6000 high-frequency memory, heat dissipation is Longshen 2nd generation 360 water cooling, this configuration can be said to be no bottleneck, can give full play to the real strength of ZOTAC RTX4090PGFOC graphics card.
AI performance testing
Now that you've covered the testbed, it's time to start today's focus. AI as one of the most cutting-edge technologies of this era, now almost everyone, all industries are talking about AI, our old Huang is no exception, on GTC, on COMPUTEX2023 are talking about AI, and even saying "now is the iPhone moment of AI", in order to let everyone feel the charm of AI, further launched a series of AI products and services, shocked the audience. Without further ado, we start witnessing the AI performance of this top-of-the-line ZOTAC RTX4090PGFOC graphics card.
AI painting
When it comes to AI painting, we have to mention the recently very popular StableDiffusion, which is commonly known as AI painting, by typing in some keywords, let AI paint the picture that everyone wants in their minds, this usage can be described as quite magical, many original artists after seeing the emergence of such tools are jokingly saying that we will be replaced by AI in the future.
StableDiffusion is a text-to-image generator based on a potential diffusion model, allowing users to enter text arbitrarily to generate high-quality, high-resolution, high-fidelity images. Compared to the equally hot Midjourney tool, StableDiffusion's advantage is open source and high controllability. This is also the reason why many players choose StableDiffusion when experiencing AI painting.
Not only that, but StableDiffusion has a huge advantage, and its support for local operation means that almost anyone can use their own computer for AI painting, and the barrier to entry is extremely low. Here we will use ZOTAC RTX4090PGFOC for testing, in order to facilitate the knowledge of its real performance, we will add other graphics cards in the follow-up test for your comparison reference.
Let's start with a simple test, using the description provided by NVIDIA to generate a 768x768 size image, the parameters of the drawing are: using the v2-1768-emapruned model; teps: 50; CFGscale: 7.5, a set of 10 drawings, 3 sets at a time. The actual measurement can see that ZOTAC RTX4090PGFOC rode the dust, 90 seconds to complete the test, converted to 3 seconds to draw a picture, this speed I believe that many artists are nervous.
After reading the hors d'oeuvres, let's draw some little sisters that the audience loves, and we used a majicmixRealistic_v6 model, with FilmVelvia2Lora and a plug-in that fixes eyes and postures. Steps: 50, Collector: Euler, CFG: 7.5, Seed: 172450070, Size: 1024x768, a total of 1 group, 6 images.
There is no doubt that the ZOTAC RTX4090PGFOC is still a ceiling-like existence, with a completion speed of 35 seconds, compared to the second-place RTX4080, it already has a 10-second advantage, not to mention other graphics cards. In other words, if you're like me and love AI painting, choosing a ZOTAC RTX4090PGFOC will satisfy all your imaginations.
AI HD photos
After watching AI paintings, you may have been impressed by the amazing performance of AI. In fact, AI can do much more than drawing, it can also shine in the field of processing photos. At this stage, it is relatively common for AI to enlarge photos non-destructively. With AI, you can enlarge otherwise low-resolution photos to larger sizes, while increasing resolution. This is a lifesaver for many old photos.
However, AI is not brainless to increase the resolution, but uses the function of AI deep learning to guess and fill in the content of the picture, bringing more details to the picture. Somewhat similar to the complement in DLSS we talked about earlier, AI's educated guesses and complements can bring your old photos to life.
In the AI application software ON1ResizeAI2023, we use AI to increase the resolution of multiple pictures by 200%, at this time the speed of ZOTAC RTX4090PGFOC is 6 seconds, and the previous generation of Kahuang RTX3090Ti is 9 seconds, in fact, you don't think that this is a 3-second gap, if your engineering amount is more complicated, then the fast 3 seconds add up to more than 30% performance improvement.
AI upscales video resolution
AI can not only high-definition photos, but also contribute a lot to improving video resolution. NVIDIA did not launch the function of improving video resolution when the RTX4090 was first released, and this epoch-making feature was officially launched until the RTX4070 series was launched later. It is the RTXVideo SuperResolution video super-resolution technology (RTXVSR), known as the video version of DLSS. Now there is just a ZOTAC RTX4090PGFOC graphics card, so use this graphics card to see if the top performance and super AI technology can improve the low-resolution video to unprecedented effects.
First of all, to turn on RTXVSR you need your graphics card to be 30 series or 40 series, find the video image settings in the NVIDIA control panel, manually turn on RTXVideoenhancement to enable RTXVSR technology, which is divided into 1-4 levels of adjustment, the larger the number, the better the quality.
In the actual measurement link, it can be seen that after turning on RTXVSR technology, the character lines in the original video have become very clear and sharp from the original blurry, and even some details have exceeded the native 1080P, and the low-quality video also has a lot of noise, and after turning on VSR4 gear, the noise almost disappears.
In addition, because it is AI intelligence to increase the resolution, the font is a little faint in the native 480P video, and even raw edges can be seen on the edge of the font, but with the blessing of VSR, the font is sharp and visible, not losing the effect of native 1080P at all, I have to say that VSR does have something.
Here we also prepare a comparison video for you, from the video you can see the effect of VSR is obvious, the highest level VSR4 compared with the native 480P is a glance difference, after turning on RTXVSR technology, some image quality has been significantly improved, not only aliasing and noise almost all disappeared, color transitions are also smoother, the current VSR can have this performance is amazing.
Left: Native 480P Middle: VSR4 Right: Native 1080P
Of course, in addition to being able to use this function in the browser, the local browser can also enjoy the charm of black technology, VLC is the first local player that supports RTXVSR technology, and can support a variety of video formats to play, and the old videos on your hard disk can also have clear picture quality!
We will local 360P quality video RTXVSR4 rendering, from the screen, after turning on VSR technology, the rendering video on the right is indeed much clearer than the native video on the left, and the noise visible to the naked eye is reduced, and the look and feel is significantly better than the native 360P video.
If you want to know the effect of local VSR, there are also videos here to see the gap between the two, and it is indeed obvious that with the blessing of VSR, many details are clearly visible again, and some low-resolution videos are indeed improved.
Left: Local native 360P Right: VSR-4 gear
Overall, the RTXVSR technology on the ZOTAC RTX4090PGFOC is quite good, with both a good look and feel and the right power consumption, and with this new AI technology, it will bring a revolutionary new experience for players and video viewers!
AI target tracking
If you're a video post, the features described next will definitely be useful to you. In the latest DavinciResolveStudio 18, it has supported the use of graphics card acceleration function, with it you can use AI to identify and track targets in the video, in the past, if you want to use manual to do, you need to cut out the target frame by frame, not only inefficient, but also quite cumbersome to do, but now with the help of AI, it has become much simpler, can greatly improve your work efficiency, and the effect is much better than manual cutout.
When we tested with the da Vinci AIACCELERATEDMAGICMASK, we found that under the same project, the ZOTAC RTX4090PGFOC with the architectural innovation has a faster rendering time, originally RTX3090Ti took 27 seconds to complete the operation, now ZOTAC RTX4090PGFOC only takes 17 seconds! With efficiencies increased by nearly 50%, the replacement of ZOTAC RTX4090PGFOC can indeed bring greater efficiency to the later team.
Here is also to show you the actual effect, you can see that after using AI target tracking, the edges of the characters are sharp, and the movements are coherent, which is really time-saving and labor-saving than manual.
summary
Just looking at the test results, you must have been shocked by the ZOTAC GeForceRTX4090PGFOC. Although this is a consumer-grade graphics card for positioning games, its AI performance should not be underestimated. Operations such as AI painting, AI tracking targets, high-definition pictures and videos have good acceleration effects. Of course, you have to take it and the professional AI acceleration card is still far behind, but as an individual user, the current acceleration effect can already allow you to improve your efficiency.
Of course, the reason why ZOTAC GeForceRTX4090PGFOC can burst out such powerful AI performance is inseparable from its luxurious configuration, the comprehensive AdaLovelace architecture under TensorCore can shine, 24GB GDDR6X memory, providing a sufficient stage for AI to play. The leap in AI performance has also become the biggest highlight of the ZOTAC GeForce RTX4090PGFOC graphics card.
However, the RTX40 series graphics card can shine in AIGC, which is Lao Huang's foresight. On the one hand, because NVIDIA occupies the vast majority of the market share, on the other hand, NVIDIA has been deeply cultivating the AI market for many years, and the strategy of accelerating AI through GPUs is laid out from top to bottom. This is very similar to today's CUDA, advanced deployment, let their own ecology be perfected, then players will naturally choose your product.
I have to say that at this stage you want a graphics card that can take you to the AI frontier, then ZOTAC GeForceRTX4090PGFOC should be your best choice. On the one hand, it has strong performance, on the other hand, it also has a complete software ecological support, and its practicality is better than other graphics cards. Top graphics, top performance, unparalleled creative potential.