NVIDIA officially released the GeForce RTX 4090 GPU that gamers are looking forward to on September 20. It is equipped with the fourth-generation Tensor Cores and the third-generation RT Cores. The performance and ray tracing performance of AI can be improved by 2 to 4 times compared to the previous generation, and supports the new DLSS 3 technology. The smooth running of ray tracing game masterpieces at 8K resolution is also one of its highlights. As an important partner of NVIDIA, Gigabyte also launched flagship graphics card products based on RTX 4090 GPU at the first time, and the GeForce RTX 4090 GAMING OC 24G Magic Eagle was the first to arrive in our evaluation room. So why is the RTX 4090 so powerful? What evolutions have been made in the new NVIDIA Ada Lovelace GPU architecture it adopts? Let’s see our comprehensive analysis and evaluation.
NVIDIA Ada Lovelace architecture brings revolutionary evolution
Ada Lovelace (1815~1852) is a British mathematician and founder of computer program . He is known as the world's first programmer . So, what kind of revolutionary upgrade does the NVIDIA Ada Lovelace GPU architecture named after her bring?
process leads, GPU scale and frequency are comprehensively improved
Complete Ada AD102 GPU contains 12 GPCs, a total of 18432 CUDA units, 144 RT Cores (third generation), 576 Tensor Cores (fourth generation) and 576 texture units. The number of transistor is as high as 76 billion, far exceeding the 28 billion NVIDIA Ampere. In one GPC of the Ada GPU, there are 6 TPCs, each TPC contains 1 polygon engine and two SM units, each SM unit contains 128 CUDA units, 1 third-generation RT Core and 4 fourth-generation Tensor Cores.
Thanks to the customized TSMC 4N manufacturing process, the scale of the NVIDIA Ada Lovelace GPU architecture has reached unprecedented strength, and has also made a huge improvement in operating frequency. Taking the current flagship GeForce RTX 40 as an example, it includes 11 GPCs, a total of 16,384 CUDA units, 128 RT Cores (third generation), 512 Tensor Cores (fourth generation) and 512 texture units. The default acceleration frequency is 2520MHz, which is far beyond the 1860MHz of the previous generation RTX 3090 Ti, and it can even be higher in games. In short, the comprehensive improvement of scale and frequency has laid a solid physical foundation for the powerful performance of the NVIDIA Ada Lovelace GPU architecture.
Shader Executing Reordering (SER) Shader Execution Reordering (SER) Shader Execution Reordering (SER)
According to NVIDIA official data, the shader data throughput of NVIDIA Ada Lovelace GPU architecture can reach up to 90 TFlops (GeForce RTX 4090 is 83 TFLOPS), while the previous generation NVIDIA Amere was about 40 TFLOPS, which of course is related to the significant increase in the scale of the new generation of GPUs.
However, in addition to the surge in scale, another evolution of the NVIDIA Ada Lovelace GPU architecture is to support Shader Executing Reordering (SER) shaders to perform reordering. We know that GPUs are most efficient when completing a lot of similar work. As the ray tracing effect becomes more and more complex in the game, the shader may face a large number of divergent and disordered inefficient work tasks, which affects the rendering efficiency. With the emergence of SER technology, these messy work tasks can be classified and dynamically reorganized into more efficient tasks, thereby improving the execution efficiency of shaders. According to NVIDIA's official data, SER technology will bring about 2 times the shader performance improvement, and in actual tests of ray tracing games like "Cyberpunk 2077", SER also brings about 44% of the frame rate improvement.
The third generation RT Cores
From RTX 20 to RTX 30 to RTX 40, RT Cores has also evolved to the third generation. Simply looking at effective ray tracing computing power, the third-generation RT Cores of the NVIDIA Ada Lovelace GPU architecture has a throughput of 191 TFLOPS, which is 2.8 times that of the previous generation.
In addition, we know that RT Core in NVIDIA Turing and Ampere GPUs include acceleration units for accelerating the traversal of boundary volume hierarchy (BVH) data structures and performing intersecting test calculations of ray triangles and ray bounding boxes.In Ampere's RT Cores, BVH traversal is accelerated by Box Intersection Engine, and the ray triangle intersection test is accelerated by the triangle intersection engine, which has proven to be the most efficient engine to handle ray tracing workloads to date. In the third-generation RT Cores of Ada GPU, in addition to the above two functions, they also introduce "micromap transparency engine" (OMM) and "micromesh displacement engine" (DMM). The function of the micromapping transparency engine is to optimize fiber tracking and rendering, greatly reducing the workload of the shader, which is a great improvement in handling the ray tracing characteristics of complex objects. The specific principle is to bake the ray tracing characteristics into opaque masks, so that irregular shapes and translucent objects can be rendered faster and more accurately.
microgrid displacement engine can simplify complex graphics with a large number of faces and create a simpler model, thereby achieving the purpose of rendering complex geometric figures with basic triangles, greatly reducing the amount of computing and the use of video memory. According to official data, the micro-grid displacement engine can increase the BVH speed of ray tracing by up to 15 times, and the video memory usage can be reduced to 1/20 of the original. In short, the more complex the model, the more outstanding the optimization effect of the micro-grid displacement engine.
Fourth-generation Tensor Cores
On the Ada GPU, Tensor Cores has evolved to the fourth generation, and this time the Tensor Cores upgrade is particularly important. It is equipped with a brand new FP8 engine, with tensor processing performance up to 1.32 PetaFlops (up to five times higher than the previous generation) - note that the unit is PetaFlops in tensile billions, while we used TFlops in the previous one. Of course, the most revolutionary progress of the fourth-generation Tensor Cores is the newly added Optical Flow Accelerator, which is the core of NVIDIA DLSS3 black technology. The DLSS3 chapter will be introduced in detail later.
DLSS3
DLSS technology is no stranger to everyone. As a special black technology for improving picture quality and frame rate in RTX series graphics cards, the RTX 40 generation has evolved directly from the previous version 2.3 to the 3.0 version. The span is so large that its improvement is revolutionary. NVIDIA even claims that it represents the "new era of neural network rendering."
is different from the previous DLSS. DLSS3 further introduces optical multi-frame generation technology based on the original super-resolution function. Simply put, it is possible to generate new frames through AI calculations, rather than just generating pixels like before. The frame generation work is done by the optical flow accelerator added to the fourth generation of Tensor Cores, which of course means that the "frame generation" function in DLSS3 is exclusive to RTX 40. The optical flow accelerator can calculate the optical flow field between two frames based on DLSS2, capture the direction and speed of changing from the first frame to the second frame in the game screen, including pixel information such as particles, light irradiation, reflection, etc., and accurately reconstruct the picture by calculating the motion vector and optical flow. Simply put, DLSS2 does the calculation of the remaining pixels on the screen through AI, and DLSS3 can also calculate the next frame of the complete picture. In this way, DLSS3 is equivalent to completely reconstructing the pixels of the actual display screen 7/8, and it can also be said to reconstruct the pixels equivalent to 7 times the actual rendered pixels. In addition, since the "frame generation" of DLSS3 is carried out in the GPU, even if there is a bottleneck in CPU performance, the game frame rate can be improved.
According to official data, DLSS3 can bring up to 4 times the game performance improvement, making 8K and panoramic ray tracing games a reality. DLSS3 includes three parts: "optical multi-frame generation technology", "super resolution" and "NVIDIA Reflex". As long as developers integrate DLSS3, they can support DLSS2 by default. DLSS3 has been supported by game developers and game engines. More than 35 games and applications will support this technology. The first batch of games and applications will be launched within October.
NVIDIA Studio Optimization
In terms of productivity, NVIDIA GPU must be a top-notch efficient tool for designers, and the optimization of the RTX 40 series in NVIDIA Studio with the NVIDIA Studio driver will also provide content creators with better choices.Overall, RTX can currently accelerate more than 110 mainstream creative applications. In addition, the SDK provides these applications more efficient and provides exclusive features such as Optix, DLSS and Maxine. NVIDIA Studio also offers a full suite of creative applications including NVIDIA Omniverse, Broadcast, Canvas, and RTX Remix.
AV1 and AI live broadcast
We know that AV1 can provide clearer video images at the same code rate than H.264 and other encoding methods, and AV1 also supports some more advanced encoding features, such as HDR and low latency. The RTX 40 encoder (RTX 40 series with over 12GB of video memory even provides dual NVENC encoder) provides support for AV1 hardware encoding, while streaming software such as OBS and Discord will provide support for NVENC AV1 soon, allowing players to enjoy a more exquisite and smooth live video picture. In terms of decoding, NVENC of RTX 40 also inherits the fifth generation hardware decoder on RTX 30, supporting MPEG-2, VC-1, H.264 (AVCHD), H.265 (HEVC), VP8/9 and AV1's hardware decoding . In the AI live broadcast section, NVIDIA brings 3 new AI effects to Broadcast applications: improved virtual background, eye contact (which makes you look like you are watching a camera) and expression evaluation, and developers can easily integrate these features of NVIDIA Broadcast in their apps.
accelerated content creation
has been introduced earlier. The eighth generation dual encoder is built in the RTX 40 GPU with 12GB video memory and above, so it has a huge improvement in video export speed. As can be seen from the figure, the video transcoding time of RTX 4090 is reduced by more than half compared to the RTX 3090 Ti, greatly improving work efficiency. DaVinci Resolve, voukoder and cut-in will all provide support for RTX 40 dual encoder. In addition, RTX 40 has also greatly improved efficiency in AI video tools. For example, in DaVinci Resolve AI Magic Mask, the processing speed of RTX 4090 is 1.7 times that of the RTX 3090 Ti. In addition, thanks to the third-generation RT Cores, SER, DLSS 3 and NVIDIA Omniverse of the RTX 40 series, 3D designers can also create at 4K/60fps without the need for agents, greatly improving work efficiency. Game enthusiasts and developers can also directly remake classic games with enhanced graphics through Modder and add RTX effects through RTX Remix.
In general, the NVIDIA Ada Lovelace GPU created by the new TSMC 4N process has brought huge upgrades in architecture, frequency and energy efficiency ratio, providing relatively algebraic times of performance. The new RT Cores, Tensor Cores design, DLSS3 technology supported by optical flow accelerator, etc. are all revolutionary advances. It not only brings players a ray-tracing game experience that is both speed and picture-oriented (even directly entered the era of 8K games), but also brings extremely efficient productivity tools to content creative designers, which can be regarded as another milestone in the history of GPU development.
unprecedented sense of oppression! RTX 4090 Magic Eagle is domineering
As a new generation flagship, Gigabyte GeForce RTX 4090 GAMING OC 24G Magic Eagle (hereinafter referred to as RTX 4090 Magic Eagle) has a superlative sense of oppression. With a 3.5-slot radiator, it has a size of 340mm×150.2mm×75.2mm, which can be called the "big Mac" in graphics cards. Such a huge radiator also proves that the strength of RTX 4090 is unprecedentedly powerful.
RTX 4090 Magic Eagle is equipped with the family's unique wind power cooling system, with 3 110mm fans, equipped with nanographene lubricant, supports 3D start-stop, forward and reverse design, the heat sink's heat-sinking plate directly touches the GPU, is equipped with 10 composite heat pipes and provides air intake grille, which effectively improves the efficiency of graphics card cooling , and fully ensures that the RTX 4090 GPU can stably perform wild performance.
graphics card is equipped with a full-size metal backplate that has been enhanced with shape. The curved edge design greatly enhances the strength of the overall architecture, perfectly ensuring the resistance of the overweight graphics card to physical deformation. With the bracket included with the graphics card, there is no worries when using it in a vertical chassis.
In order to meet the different needs of players, the graphics card also provides dual BIOS mode, where players can switch the powerful OC mode or the silent mode with lower noise with one click.The auxiliary power supply part, in addition to the new 16pin power supply interface standard (the graphics card comes with a 16Pin to 8Pin×4 power cord, which can also be used in traditional power supply), is also equipped with a power supply indicator, which will flash when the power supply is abnormal, helping players quickly troubleshoot problems.
personalized part, the graphics card is equipped with Gigabyte's unique RGB FUSION lighting system. The "three-ring lighting effect-RGB fan" of the RGB lights when rotated is particularly eye-catching, and can also achieve lighting effects linkage with other accessories that support RGB FUSION to create a cool lighting effect MOD host. The graphics card is equipped with 3 DP 1.4 and 1 HDMI 2.1 interface, which can support up to 4 screen output, fully meeting the needs of players to connect to multiple display devices.
Overall, the RTX 4090 Magic Eagle has reached the level of current top flagships in terms of power supply, cooling and personalized design, and is a good choice for ultimate enthusiasts to try out the RTX 4090. Next, let’s enter the practical test section that everyone is most concerned about!
RTX 4090 ray tracing performance has increased dramatically, DLSS3 is incredible
test platform
graphics card: Gigabyte GeForce RTX 4090 GAMING OC 24G Magic Eagle
Gigabyte GeForce RTX 3090 Ti GAMING OC 24G Magic Eagle
Processor: Core i9 12900K
Motherboard: Gigabyte Z690 AORUS MASTER
Memory: Gigabyte AORUS DDR5 6000 16GB×2
Hard disk: Gigabyte AORUS NVMe Gen4 Titanium Engraving 2TB
Power supply: Gigabyte UD1000GM PG5
Operating system: Windows 11 Professional Edition
Practical performance test
benchmark performance test part, the improvement of RTX 4090 compared to RTX 3090 Ti is very exaggerated. In the FireStike test, which represents the performance of DX11, the leading amplitude was over 80%. Moreover, with the higher the resolution, the greater the leading amplitude. It can be seen that subsystems other than GPUs have gradually shown a bottleneck effect in the face of the powerful performance of RTX 4090. In the DX12 part, the RTX 4090's leading range is also close to 80%. In the DXR ray tracing part, RTX 4090 has more than doubled the RTX 3090 Ti. After all, the advantages of 128 third-generation ray tracing units against 84 second-generation ray tracing units are self-evident. Of course, in PortRoyal, RTX 4090 is also more than 80%. The actual test part of
gaming. For flagship graphics cards such as RTX 4090, the pressure on the GPU using 1080P resolution is too small, which will only highlight the bottleneck of subsystems such as processors. Therefore, we only test at 2K and 4K resolutions. First of all, there are 2K test parts. Among the 13 games, 11 of them have obvious improvements, and the advantages of "Doom: Eternal" even reach more than 58%. Online games like "CS:GO" and "World of Warcraft: Shadowlands" do not have enough pressure on the GPU, and obviously cannot exert the power of RTX 4090, and the performance of the processor has also become a bottleneck here. Overall, the average advantage of RTX 4090 over RTX 3090 Ti at 2K resolution is about 130%. It is obvious that the pressure of 2K resolution seems too weak in front of the RTX 4090, and the GPU cannot be fully exerted due to the limitations of subsystems such as processors. So let's take a look at the situation at 4K resolution.
As the resolution increases to 4K, the performance advantages of RTX 4090 are becoming more and more obvious. In "Killer 3", RTX 4090 is even more than 80% ahead of the RTX 3090 Ti. This time, even games like "World of Warcraft: Shadowlands" that are seriously restricted by the processor bottleneck have also achieved a frame rate increase of more than 28% with the support of the powerful performance of RTX 4090. Overall, the average frame rate of RTX 4090's 4K games leads the RTX 3090 Ti about 154%. However, it is obvious that ordinary 4K games have not completely squeezed out the true strength of RTX 4090, and the GPU pressure is still insufficient. Then, it’s time to start the ray chase game!
For players who need extreme ray tracing images and super high e-sports level frame rates, 2K resolution is a relatively balanced setting, and the pressure for GPUs like RTX 4090 is only moderate. Judging from the test situation, except for some games that have a smaller improvement, other games have basically increased by more than 50% compared to the RTX 3090 Ti, and can be increased by more than 80%. Overall, the average increase is about 57%, which is also very considerable.Of course, this is not enough. Let’s take a look at the most extreme 4K ray tracing squeeze!
4K ray tracing game part, we also added the DLSS test. RTX 4090 has more fourth-generation Tensor Cores, which can just show its skills. Sure enough, under the super high load press of 4K+ extreme ray tracing, the advantages of RTX 4090 over the RTX 3090 Ti have been improved again. In "F1 2022" and "Having Light 2", RTX 4090 really achieves twice the performance of the RTX 3090 Ti! In many other games, RTX 4090 also maintains an advantage of 70% to 80%. In summary, whether DLSS is turned on or off, the average advantage of RTX 4090 remains at around 170%. At the same time, we can also see that after turning on DLSS, the RTX 4090 has achieved huge performance improvements in many games, allowing 4K ultimate ray tracing games to run at smooth frame rates of more than 100fps or even 200fps.
In short, judging from the current situation of popular games, only under the 4K ultimate ray tracing image quality settings can the powerful performance of RTX 4090 be tried, and from the frame rate of the test game, it is really necessary for players to RTX 4090 to have a 4K/144Hz e-sports monitor . The previous technical analysis part of
has been introduced. The RTX 40 series also has added an optical flow accelerator. In conjunction with the DLSS3 technology released at the same time, it can generate complete picture frames through AI calculations, thereby greatly improving the frame rate and reducing frame delay. Currently, 35 games and applications have announced that they will support DLSS3 technology soon, and we selected several of them for this test. In games that support DLSS3, switch options are provided for DLSS super resolution, frame generation, and NVIDIA Reflex. These are actually the three components of DLSS3, and turning off frame generation is equivalent to using DLSS2.
From the test results, we can see that in "Cyberpunk 2077", with the support of DLSS3 frame generation, the average frame rate of RTX 4090 even exceeds 2.3 times that of RTX 3090 Ti! The DLSS3 beta version of " Against the Water Cold " is even more exaggerated. The ray tracing effect of the test copy specially designed for DLSS3 is close to the movie screen, and the pressure on the GPU is abnormal. Even the RTX 4090 only has 17fps in the 4K without DLSS. After turning on the DLSS3 performance mode, the frame rate has been directly increased by nearly 4 times to 84fps. The RTX 3090 Ti without an optical flow accelerator has only 25fps in the 4K+DLSS performance mode, and the RTX 4090 is 236% ahead! It can be seen that the more ray tracing special effects in the game, the more complex it is, and the greater the performance improvement brought by DLSS3. Therefore, facing the future panoramic ray tracing masterpieces, the RTX 40 series with optical flow accelerator supports DLSS3 is obviously a must-have. "Lyra" has relatively little pressure on graphics cards, and after turning on DLSS3 and frame generation, the RTX 4090 also has a frame rate advantage of up to about 65% compared to the RTX 3090 Ti. DMark has also released an internal beta version that supports DLSS3. It can be seen that the frame rate of RTX 4090 has increased by 200% after turning on DLSS3, and the RTX 3090 Ti supported by DLSS2 is also 117%. It can be seen that the optical flow accelerator added to the RTX 40 series is indeed a very powerful design.
We also briefly compared the game picture quality supported by DLSS3. As can be seen from the figure, after AI machine learning, DLSS has been very good at optimizing image quality. Even for the performance-oriented DLSS performance mode, the sharpness and details of the picture are even better than turning off the native picture of DLSS (pay attention to the texture of the taillights). This is why NVIDIA also recommends that everyone choose the DLSS3 performance mode at 4K resolution, and it has achieved both image quality and high frame rate .
For designer users, NVIDIA Studio is too familiar. So how powerful is the RTX 4090 with 24GB ultra-large video memory as a productivity tool? We chose mainstream 3D rendering tools and video editing software to conduct actual testing. From the test results, we can see that the rendering performance of RTX 4090 compared to the RTX 3090 Ti has almost doubled, and most projects have a performance improvement of 80% to 100%. As for the video transcoding part, since the RTX 4090 is supported by dual encoder, the transcoding speed in the clipping is almost twice that of the RTX 3090 Ti. In addition, the RTX 4090 also provides AV1 encoding support, and the RTX 3090 Ti does not have a corresponding hardware encoder, of course, it cannot be supported.
In short, for 3D designers or video effects and editing workers, upgrading from RTX 3090 Ti to RTX 4090 is basically equivalent to achieving twice the processing efficiency, which is obviously very worth it.
power consumption temperature and overclocking test
RTX 40 series adopts the TSMC 4N manufacturing process, and the frequency and energy efficiency are greatly improved compared to the previous generation, which also makes its GPU frequency very expected to be increased to above 3 GHz. Of course, the Gigabyte RTX 4090 Magic Eagle is also equipped with a powerful 10 heat pipe 3 fan 3.5-slot radiator, with the purpose of which is to allow the RTX 4090 to release power stably and perform performance. Judging from the GPU-Z information, the default GPU acceleration frequency of the RTX 4090 Magic Eagle is 2535 MHz (up to 2775 MHz in the game), and the power wall can be adjusted upward by 33.3%, which shows that there is still a lot of room for improvement.
FurMark actual test frequency is 2745 MHz, the default test power is up to 441W, the GPU temperature is up to 63.1℃, the junction temperature is 71.9℃, and the heat dissipation performance is also quite good. Since the radiator is so powerful and the power limit is so high, of course we have to try overclocking . Through repeated attempts, without adjusting the voltage, the GPU frequency of the RTX 4090 Magic Eagle successfully broke through the 3 GHz mark. The measured frequency in 3DMark can reach 3015 MHz. At this time, the TGP also rushed to more than 500W. The FireStrike Ultra score increased from 25082 to 26527 at the default frequency, and the working temperature increased slightly. It can be seen that its power supply and heat dissipation design can indeed fully control the monster-level GPU of RTX 4090.
Summary: Enjoy the ultimate artifact of ray tracing games, RTX 4090 Demon Eagle is the choice of supreme
The newly created NVIDIA Ada Lovelace GPU architecture has brought revolutionary progress to the RTX 40 series. In addition to the super-large chip scale and high frequency gain obtained from the TSMC 4N process, the third-generation RT Cores and the fourth-generation Tensor Cores also shines, especially the newly added optical flow accelerator, which allows the AI frame generation of DLSS3 technology to be achieved, thereby greatly increasing the game frame rate, and ultimately bringing 2 to 4 times the performance improvement of ray tracing game, making it easy to run smoothly at the e-sports level frame rate, and also allows players to calmly deal with the panoramic ray tracing game and 8K ultra-clear game with more realistic pictures in the future. In addition, from the previous tests, we can also see that the RTX 4090 has doubled performance growth compared to the previous generation in terms of 3D rendering output and video editing. It is really a top-level efficient tool that must be bought for the productivity application of "time is money".
Therefore, for enthusiastic players who pursue the ultimate gaming experience and designer users who pursue top efficiency, RTX 4090 is undisputedly the strongest choice at present, and it is necessary to start with one step. Of course, in the super public version of RTX 4090, Gigabyte RTX 4090 not only has luxurious materials and excellent heat dissipation design, but also has a power release space of up to 133.3%. It is a luxury flagship tailor-made for DIY players, game enthusiasts and high-performance users, and is worthy of strong recommendation.
