Nvidia GeForce RTX 4090 review: a luxury monolith that bets everything on raytracing and 4K


Nvidia uses an AD102 chip based on the Ada Lovelace architecture.  It is engraved in 5 nm by TSMC.

Nvidia uses an AD102 chip based on the Ada Lovelace architecture. It is engraved in 5 nm by TSMC.

© Nvidia

The GeForce RTX 4090 is Nvidia’s first graphics card to use the Ada Lovelace architecture. In absolute terms, this architecture does not radically change the way Nvidia’s GPUs work. On the contrary, we are even more in line with what the Ampère architecture proposed. Thus, the organization of the different blocks does not differ. However, the focus has been on their effectiveness in raytracing and artificial intelligence. At the same time, the manufacturer dangles the imminent arrival of games (or game updates) exploiting new raytracing effects, allowing to approach a little more the chimera of photorealism. These new effects will be heavy to manage for the GPUs and justify the efforts undertaken by Nvidia on the Ada architecture.

Graphical representation of the AD102 GPU with all its active units.  A few blocks of the GeForce RTX 4090 are disabled and leave room for a hypothetical GeForce RTX 4090 Ti.

Graphical representation of the AD102 GPU with all its active units. A few blocks of the GeForce RTX 4090 are disabled and leave room for a hypothetical GeForce RTX 4090 Ti.

© Nvidia

The SM calculation blocks thus integrate the 4th generation of Tensor Cores units in order to accelerate processing on artificial intelligence. These units are said to be up to twice as fast as 3rd generation units. The 3rd generation RT Cores are also advertised as twice as fast as the previous ones, but above all are up to 10x faster for processing microtriangles (primitives that allow more detailed rendering).

The Nvidia SER can significantly improve the order of data processing.

The Nvidia SER can significantly improve the order of data processing.

© Nvidia

Nvidia also took the opportunity to review the way in which the data to be calculated are processed. The Shader Execution Reordering (SER) thus makes it possible to analyze and reorganize the data so as to gain in efficiency on the rendering in raytracing, which can result in a gain in throughput of up to 44%.

AD102: a chip with 76.3 billion transistors

The circuit board retains the original V-shaped appearance. The power stage is slightly beefier than that of the GeForce RTX 3090 Ti.

The circuit board retains the original V-shaped appearance. The power stage is slightly beefier than that of the GeForce RTX 3090 Ti.

© Nvidia

On the GeForce RTX 4090, Nvidia uses a large AD102 GPU. This does not exploit all the possibilities in terms of number of units, which leaves room for a hypothetical GeForce RTX 4090 Ti. The chip is engraved on the 4N process of TSMC and not N4. A nuance that is very important: the 4N used here is a 5 nm engraving process, optimized for Nvidia’s needs – therefore not 4 nm.

Still, the GPU has no less than 76.3 billion transistors, almost three times more than the GeForce RTX 3090 chip. 608.5 mm², a slightly smaller area than Ampère (628.4 mm²).

Technical characteristics of the GeForce RTX 4080 and GeForce RTX 4090.

Technical characteristics of the GeForce RTX 4080 and GeForce RTX 4090.

© Nvidia

With 11 GPC compute blocks for 64 TPC sub-blocks, the GeForce RTX 4090 has a high number of compute units. Generational growth is strong, with 16,384 CUDA units, 128 RT Cores units, and 512 Tensor Cores units, which is an overall increase of 52% across the board over the GeForce RTX 3090 Ti. Incidentally, we note that Nvidia has inflated the different cache levels, the L1 going from 10 MB to 16 MB, the L2 from 6 MB to 74 MB and the file register from 22 MB to 32 MB.

And even if the number of calculation units is high, the frequency jumps a big step to reach 2520 MHz announced in GPU Boost, an average frequency that we could observe in games. More than 700 MHz gained over a generation, that’s interesting.

The graphics memory type is unchanged and remains GDDR6X. The 24 GB of memory make it possible to anticipate the most distant needs in the video game field and above all lend the card well to creative use. The frequency does not change compared to the GeForce RTX 3090 Ti, nor does the interface. Thus with 2625 MHz and a 384-bit bus, we are again faced with a general bandwidth of 1 GB/s.

450 watts to cool

The Founders Edition model is massive and weighs 2.2 kg.

The Founders Edition model is massive and weighs 2.2 kg.

The energy envelope is set at 450 watts, a value significantly higher than that of the GeForce RTX 3090 (350 watts), but similar to that of the GeForce RTX 3090 Ti. A very high value for a card intended for video games, even though we are talking about an elitist model.

The cooling system

The “dolphin” cooling system is retained on this model.

© The Digital

Nvidia, on the other hand, has absolutely not skimped on the means to properly cool its graphics card. Visually, the cooling system is quite similar to that used on the GeForce RTX 30. We thus find the double “dolphin” fan system, sucking in the air from below the card and expelling it above, towards the PC processor. The finish is always exemplary and only the curved edges make it possible to differentiate this model from the previous one.

In addition to this aesthetic detail, some other less visible points have been reworked. The fans gain in diameter (from 110 to 116 mm) and leave aside the ball bearings in favor of an oil bath; graphics card height reduced from 2.7 slots to 3 slots — due to thicker heatsink; the vapor chamber has better heat conductivity and the number of heat pipes has been increased.

GeForce RTX 4090 and RTX 3090 look alike.  The 4090 thickens and has some curves on its edges and larger fans.

GeForce RTX 4090 and RTX 3090 look alike. The 4090 thickens and has some curves on its edges and larger fans.

© The Digital

All this results in a beautiful baby of no less than 2.2 kg that should be fixed firmly to the case to avoid any tearing of the PCIe 4 port on which it will be fixed – because yes, the interface remains in PCIe 4 and only the card’s external power supply changes to the PCIe 5 standard. A power adapter is provided and allows the use of three (or four, if overclocking) classic 8-pin PCIe cables instead of the PCIe 5 cable.

The reference card – called Founders Edition – also offers four video outputs: three DisplayPort 1.4a and one HDMI 2.1a. The latter is able to output an 8K HDR signal at 60Hz or 4K HDR at 120Hz and also supports variable refresh.

The GeForce RTX 4090 makes an interesting generational leap, but one that comes at a high price.

The GeForce RTX 4090 makes an interesting generational leap, but one that comes at a high price.

© The Digital

The GeForce RTX 4090 is undeniably an excellent graphics card, capable of running any game in 4K at high speed, with a very high level of manufacturing quality and operating in complete discretion. With the Ada Lovelace architecture, Nvidia also wanted to project itself into the future which it sees as full of raytracing effects – we will have to see if the developers share this vision, but we must recognize that Nvidia has always been able to show itself persuasive with them.

It is also on raytracing that this model stands out with technical characteristics that allow it to take off compared to the previous generation of GeForce. This very significant leap forward in RT performance is accompanied by the arrival of DLSS 3 and its image insertion system. A key feature that should allow the GeForce RTX 4090 to run next-gen games well with significantly greater ray effects.

The cooling system of the GeForce RTX 4090 is devilishly efficient and more discreet than that of the GeForce RTX 3090 (in the background in the photo).

The cooling system of the GeForce RTX 4090 is devilishly efficient and more discreet than that of the GeForce RTX 3090 (in the background in the photo).

© The Digital

Still, this nice technological package has a cost. First of all, an energy cost, with an electricity consumption that can climb to 450 watts. Everyone will have their own opinion on the interest of offering consumer graphics cards that consume so much in a period when energy sobriety is on everyone’s lips. This choice also affects the power supply of the PC, an 850 watt model being for us the bare minimum – with processors now drawing more than 200 watts.

Finally, an economic cost, with a price of 1949 €, a sum that will keep the GeForce RTX 4090 away from a significant share of the stock market – not to mention the price of a possible change of power supply. Inflation therefore won over Nvidia, which had, as a reminder, launched its GeForce RTX 3090 at €1,549. It remains to be seen how the Radeon RX 7000s will fare, both in terms of performance and value for money.


#Nvidia #GeForce #RTX #review #luxury #monolith #bets #raytracing

By admin

Leave a Reply

Your email address will not be published. Required fields are marked *