Learn more about the Ada Lovelace architecture and DLSS 3.0

The previous generation, called Ampere, was rightly called a revolution by many. NVIDIA was one of the first to tie neural networks and machine learning algorithms to its hardware, and the result of these experiments was a chic thing called Deep Learning Super Sampling (or simply DLSS). This is the so-called supersampling algorithm, which upscales the image in the style of modern TV processors, at the same time increases the number of FPS and increases the clarity and resolution of the picture. When developing Ada Lovelace, NVIDIA engineers focused on refining existing technologies.

In particular, the third version of DLSS was presented with a built-in NVIDIA Reflex optimizer, which, according to NVIDIA representatives, can artificially increase the frame rate by 2-4 (!) times. Unlike older versions, DLSS 3.0 does not just draw pixels in existing frames, but also independently generates entire intermediate frames. To do this, DLSS was taught to work with four types of data (frames, optical flow, motion vectors and depth), which increased the accuracy and efficiency of the engine. Then, the algorithm built into DLSS 3 analyzes all four types of data and selects the most suitable for each particular frame.

DLSS 3.0 performance in games.

It should be noted that the optical flow field is a completely new concept in the field of scaling, which is the responsibility of a separate hardware accelerator. Its essence boils down to finding and identifying links between the movements of pixels and objects in subsequent frames. Correlating this data with motion and dynamics data will allow you to more accurately recreate changes in the scene geometry. It is because of this feature that DLSS support was added only to new video cards. According to NVIDIA representatives, gamers with RTX 30XX cards during beta tests regularly encountered lags and artifacts when activating DLSS 3.0, so they decided to remove this option. For a while or forever, even NVIDIA does not know yet.

The work of DLSS 3.0 was demonstrated using the example of Cyberpunk 2077 and Microsoft Flight Simulator. In the first case, the frame rate was increased by 4 times, in the second - by 2. This is due to the fact that the performance of Flight Simulator is highly dependent on the processor. But what is generally incomprehensible, why does Cyberpunk 2077 produce 22 FPS on the flagship video card of 2022? There is a suspicion that NVIDIA turned on the laggy overdrive raytracing mode to make the comparisons look more impressive.

MSI GeForce RTX 4080 12GB VENTUS 3X OC 

During the announcement of the new generation, the guys from NVIDIA decided to cheat a little, using their experience in developing mobile video cards. Therefore, some important technical details hit the press only a day after the announcement. Initially, it was known that the flagship RTX 4090 is based on the AD102 graphics chip. And that was the assumption that both versions of the RTX 4080 are based on the same GPU, right? So right?

As it turned out, no. We follow our hands. The RTX 4080 12GB version includes 7680 CUDA cores and a 192-bit bus. In turn, the RTX 4080 16GB carries 9728 CUDA cores on board, and the bus width has been increased to 256-bit. That is, in fact, these are, as it were, different video cards with the same last name. And if you look under the hood, they are based on different GPUs. The 16 GB version is based on the AD103 chip, while the 12 GB version is based on the previous generation GA104 chip that was used in the RTX 3070.

In order to hide the cheating, NVIDIA slightly raised the clock speeds of the 12-gigabyte version so that it looked more advantageous against the background of its older brother. So in normal mode, it operates at a frequency of 2.31 GHz, and if necessary, it can automatically overclock to 2.61 GHz. In comparison, the RTX 4080 16 GB operates at frequencies ranging from 2.21 to 2.51 GHz, but it has a wider different core, a wider bus and more CUDA cores, so the performance is higher. But if you don’t know this and you’re accustomed to looking only at frequencies, then it may seem that the younger model is even more powerful. While it's actually more fair to call the 12GB version RTX 4070 and not be disingenuous, it would be hard to sell for $900 if that's the case.

MSI GeForce RTX 4080 16GB VENTUS 3X OC 
+7 photo

The RTX 4080 with 16 GB of VRAM belongs to the middle level, bypassing the younger version of the RTX 4080 with 12 GB in many ways, but at the same time significantly inferior to the flagship RTX 3090 with 24 GB of G6X VRAM. During production, NVIDIA engineers used new streaming multiprocessors and fourth-generation tensor cores, which doubled the efficiency of the built-in supersampling algorithms.

Despite the new architecture (Ada Lovelace, TSMC 5 nm manufacturing process), in terms of design, the RTX 4080 largely repeats its predecessor, the RTX 3080. it provides for separate third-generation RT cores. But it uses a brand new AD103 die with 9728 CUDA cores, a GDDR6X buffer and a bandwidth of 720 GB / s as the basis. According to the statement, all this goodness turns out to be 2-3 times faster than the predecessor RTX 3080 Ti. The recommended cost of the model is $1199.

The RTX 4080 16GB version by MSI technically corresponds to the reference sample, but everything else has been redone. In particular, the standard cooling system was replaced with a radiator with three fourth-generation Torx Fans, the design was reinforced with metal backplates, and all the lighting was sent under the knife so as not to confuse gamers with price tags.

MSI GeForce RTX 4090 GAMING X TRIO 24G 
+6 photo
+1 video

The RTX 4090 is a flagship graphics card that delivers incredible performance by 2022 standards. It uses the new AD102 die with 16,384 CUDA cores, and support for a 384-bit memory bus. The volume of the video buffer reaches an impressive 24 GB of GDDR6X memory with a bandwidth of 1008 GB / s.

In its production, new streaming multiprocessors, fourth-generation tensor cores and class 3.0 RT cores were used, thanks to which the efficiency of ray tracing and artificial image scaling increased significantly. The computing power of this colossus reaches 83 Tflops, and the advantage in gaming performance over the cards of the previous generation can be fourfold.

During the announcement, Jensen Huang emphasized that this is an ideal card not only for gamers, but also for content creators. Therefore, the RTX 4080 boasts NVIDIA encoders, support for the NVIDIA Broadcast application, and integration with the NVIDIA Studio graphics platform. Also, support for integrated NVIDIA Reflex technology has been sewn into the video card drivers to reduce input lag. All this resulted in a rather high recommended price - $1599.

For the GAMING TRIO modification, MSI engineers redesigned the casing and cooling system, adding a massive metal backplate and three branded fifth-generation TORX Fans to the design. In addition, the video card was decorated with LEDs with support for the MSI Mystic Light system and a dual BIOS for quick switching between settings.

Gaming performance

Comparison of performance with the previous generation of NVIDIA graphics cards.

The gaming performance of new products is definitely impressive, but so far there are no honest tests, but only official info, it is better to take it with a grain of salt. In particular, in the comments under one of the Cyberpunk demos, people literally blaze with the fact that NVDIA deliberately includes unnecessary graphics settings like RT Override in order to intentionally reduce the frame rate, and then artificially increase it using DLSS 3.0 and proudly declare, that “RTX 40XX cards are 4 times more powerful than their predecessors, run to the store. Sorry for the grumbling.

According to preliminary data, the RTX 4090 performs 1.5 to 2 times higher than the RTX 3090 Ti. The tests were carried out in Assassin's Creed Valhalla, Resident Evil Village and Microsoft Flight Simulator. In next-gen games with a focus on lighting and ray tracing, the RTX 4090 has an almost fourfold advantage over the RTX 3090 Ti. Whether this is related to DLSS 3.0 or not, NVIDIA did not specify.

The RTX 4080 with 16 GB in old-gen games without ray tracing outperforms the RTX 3090 Ti by an average of one and a half times. With ray tracing and DLSS enabled, the performance difference can be two or three times. Serious result. In fact, the real alternative to the old RTX 3090 Ti is the younger model GeForce RTX 4080 with 12 GB of video memory. In normal tests, they go head to head, and the difference in FPS rarely exceeds 10 frames. When ray tracing and DLSS 3.0 are activated, the novelty turns on the afterburner and is carried forward at twice the speed.

Other technical details

The evolution of RTX graphics cards.

After publishing all the technical specifications on the NVIDIA website, it turned out that the new products would be compatible with old power supplies. According to preliminary leaks, the next generation of RTX was supposed to switch to a new format power supply with a fifth generation PCI-E connector. In fact, it turned out that the RTX 4090 and RTX 4080 are designed to use both the old 8-pin connectors and the new PCI-E 5 connectors.

The power consumption of cards has grown, but not nearly as much as insiders feared. According to official figures, the flagship RTX 4090 with 24 GB of VRAM consumes 450 W (note: an 850 W PSU is recommended for the entire system), while the RTX 4080 requires from 285 to 320 W (depending on the amount of VRAM). And that's good news, because the RTX 30XX graphics cards used almost the same amount of power, so most people won't have to bother with a power supply upgrade. NVIDIA also mentioned that they do not exclude the appearance of special overclocking versions with a power consumption of under 600 W, but if this happens, then it will be a piece product for enthusiasts.

All new devices are equipped with modern HDMI 2.1a ports and are therefore capable of transmitting HDR video in 4K resolution with a refresh rate of 120 Hz, which fully meets the requirements of next-gen games. But the DisplayPort 2.0 interface promised by insiders was not here, instead, version DP 1.4a is used.

Afterword

The same design as the previous generation.

NVIDIA once again has a line of technically advanced video cards with serious performance and a bunch of additional goodies on board (DLSS, Ray Tracing, Reflex, etc.). And competitors will once again have to copy, simplify and catch up. The problem is that NVIDIA has been sitting on the throne for too long, has become a little bronzed and, judging by the prices, has finally lost touch with the players. Comments under the news and the first reviews are literally inundated with hate from former fans of the brand, who sincerely do not understand how a consumer, albeit very powerful, video card can cost almost $2,000, although even in pre-Soviet times, you could assemble a serious gaming PC for this money, Yes, a little more would be left on the monitor.