Nvidia announces the GeForce RTX 4090 and 4080, and explodes performance… and prices?

The new generation of “Ada” GPUs, at the controls of the new GeForce RTX 4000 series, take advantage of new engraving finesse and new technologies to explode performance. Especially in ray tracing, where DLSS 3 promises miracles.

The Ada (Lovelace) architecture, which is the heart of Nvidia’s new series of GPUs, pushes the current limits in terms of graphics calculation very far. During his annual conference, the champion of graphics chips presented the two flagships of this new generation, the RTX 4080 and RTX 4090. Very high-end that overflows with transistors, drips with technologies… and literally explodes in price.

On the design side, we’ve known it for a while now: the chips of this new generation are no longer 8-nanometer products from Samsung factories, as was the case for the RTX 3000s. The new batch now comes out of the best factories in the world, that from Taiwanese TSMC. But what does this technological change mean? Simply that with TSMC’s state-of-the-art etching (4N process, an improved 5nm), Nvidia was able to cram transistors like never before. For the “AD102-300” chip of the RTX 4090, no less than 76 billion transistors make up the chip, which makes it the densest single-block consumer processor ever launched ahead of the Apple M1 Max ( the M1 Ultra is a chip combining two M1 Max chips, i.e. 2×57=114 billion transistors).

The number of CUDA cores is literally exploding. After the 10,496 CUDA Cores of the RTX 3090, the RTX 4090 now offers 16,384. A gain of 60% which is combined with a significant increase in frequency. It goes from 1.7 GHz to 2.5 Ghz on the high-end model. Coupled with the 384-bit bus, which still manages 24 GB of GDDR6X RAM, we get a monstrous card in power. However, it remains in the footsteps of its ancestor since the RTX 4090 retains a TDP of 450 W. That’s a lot, it’s huge, but it’s an exceptional card. A card that should deploy 82.6 TFLOPS, which would make it 2.3 times more efficient than the RTX 3090 it replaces. Obviously, its price is commensurate with its power: 1,599 dollars without taxes in the United States. Understand around 1,900 – 2,000 euros including tax in our European latitudes.

Based on the same architecture and on the finesse of the engravings, two chips will equip the two versions of the RTX 4090. For the 12 GB version of memory on a 192-bit bus, the AD104 chip develops the power of its 7,680 CUDA cores in a thermal envelope of 285 W. The AD103 is a GPU that embeds 9,728 CUDA cores and 16 GB of GDDR6X, and will see its fans blowing harder to evacuate the 350 W. The first will cost 899 dollars without taxes (1,100 – 1,200 euros including tax?) and the second 1,199 dollars (1,450 – 1,550 euros including tax?).

DLSS 3: ray tracing performance in orbit

What really changes between the previous generations Turing (RTX series 2000) and Ampere (RTX 3000) and this generation Ada (RTX 4000) is the explosion in the number of “Tensor Core” that are in charge (basically) AI calculations. Compared to the previous generation, if the shaders have been multiplied by 2.25 and the RT Cores (in charge of ray tracing) by 2.56, the number of Tensor Core increases from 320 to 1400. A multiplication by 4.4 which is notably at the origin of the explosion of performance in ray tracing.

Yes, because if it is indeed the RT Cores which do the “ray tracing” calculations, it is the Tensor Cores which are in charge of improving image quality (denoising, in particular). But above all, they are the ones who drive Nvidia’s fantastic scaling technology: DLSS. The most successful of its kind, well ahead of AMD’s FSR 2.0 or Intel’s XeSS, DLSS technology makes it possible to calculate images of a game with full options (and therefore ray tracing activated) in a lower definition than that of the screen to speed up the frame rate. Then, it enlarges without any loss the frames to obtain a game in 4K quality which was not calculated in 4K.

The DLSS (Deep Learning Super Sample) comes in version 3.0. It takes advantage of the power of the RTX 4090 and 4080 to deploy a new weapon called Ada Optical Flow. It consists nothing less than using silicon “neurons” and state-of-the-art algorithms to anticipate changes in a scene. This allows, at constant image quality, to increase the number of images per second even more. The new DLSS also takes advantage of a new technology called Shader Execution Reordering (SER) which performs intelligent sorting of tasks by similarity so that they are executed simultaneously – GPUs are very good at quickly performing lots of similar tasks. It’s this SER that makes DLSS 3.0 so good.

According to Nvidia, the performance boost ranges from “a lot” to “absolutely huge”. Applied to the RTX 4090, the technology already manages to double the number of frames per second in 4K on Microsoft Flight Simulator, a game renowned for the delicacy of these extended environments. But that goes to one times four out of Cyberpunk 2077 (in a version overdrive ”, improved by Nvidia) and up to x5 on Portal RTX ! The gains are logically a little lower with the RTX 4080s, but it’s already +50% to +300% compared to the performance of the RTX 3080 Ti on flight simulator and cyberpunk ! If this performance gain cannot be applied to all games – we are talking here about an acceleration of DLSS-compatible titles and using ray tracing – we are however entitled to expect significant gains in “old” titles making simple calls rasterization.

Like any new generation of GPUs, the RTX 4000s also hide improvements not related to pure 3D, such as a new multimedia engine supporting AV1 hardware encoding or native 8K video capture (4090). But it’s obviously performance gaming which will be mainly scrutinized.

The RTX 3000 will remain in the catalog

While the prices for these three new GPUs are scary and seem straight out of your covid nightmares, the reality is that these chips aren’t replacing the RTX 3000s (yet). clarified: the generations of cards will complete each other for a while. This will allow Nvidia to take advantage of a range effect while waiting for the arrival of the classic versions – 4070, 4060 and, perhaps, 4050.

Nvidia has every interest in continuing to extend the life of its old GPUs – and that’s good, it still has plenty left – because by the end of the year, AMD will in turn present its new GPU architecture, called RDNA 3. In the cabbage five years ago, AMD managed a return to “the game thanks to RDNA and then RDNA 2, two generations of GPU architecture that have lived up to the promise of a +50% improvement in performance per Watt each time. Already on the heels of Nvidia with RDNA 2, AMD is currently still behind in ray tracing technologies. But its recent successes make it an increasingly credible threat. It will have to be to be able to compete with the new monsters which Nvidia has just given birth to.

Leave a Comment