NVIDIA is currently celebrating its GTC, but this has focused more on markets outside the PC world in which Jensen Huang’s company has become increasingly important since they will go from being a graphic hardware designer to include artificial intelligence. But what about future graphics cards like the RTX 30 SUPER or even the next generation with the RTX 40 based on the Ada Lovelace architecture?
The RTX 4090 will consume 600 W
A few days ago we told you how the new 12-pin power port that will be released in PCI Express 5.0 will allow the cards to be fed with 600 W of power, a connector that is none other than the same connector that NVIDIA had already included in some models of their RTX 30 and therefore we should not be surprised that it is used at its best in its next generation of graphics cards for gaming.
What do we know about the RTX 40? Well, little really, the most reliable sources say that they will use the TSMC 5 nm node to manufacture the chip and that it will have a total of 144 SM, an impressive figure that indicates a mammoth chip. Of course, such a level of SM also means having a communication infrastructure even larger than the current one. The consequences of it? The fact that we are going to go from the 350 W of the RTX 3090 to the 550 to 600 W of consumption of the RTX 4090.
RTX 4090 Release Date
The chip that the RTX 4090 will use as a GPU is the AD102, a beast that will surely touch the limit of the reticle as it is a large monolithic chip. Under the 7 nm node we have already seen AMD place 128 SMs with its Intinct MI200 on a single chip, so it is not within the impossible for NVIDIA to be able to place 144 units under a node that being conservative has a density between 1 , 6 and 2 times greater.
But what interests us is to know its launch date and it seems that NVIDIA has become quite traditional in this, since we will see the first models appear in the third quarter of 2022, just two years after the presentation of the RTX 30 In other words, we could see them in the market as soon as July 2022.
This means that the launch distance between the RTX 30 SUPER and the early models of the RTX 40 will be very tight. Our prediction? Rare, expensive units that Jensen Huang’s company will make to pay a premium for that additional power level.
RTX 4090 with lower performance per watt?
NVIDIA has been increasing the performance per watt of its graphics cards for several generations, this means that for the same energy consumption the newer architectures have offered higher performance. Unfortunately, if it were confirmed that NVIDIA with the RTX 40 would have decided to sacrifice the path of efficiency in its architectures for brute force, but this would make a sense and it has to do with mining, a less efficient architecture in this aspect would make it less attractive to miners.
At the moment it is still early to talk about the Lovelace architecture and that the RTX 4090 can consume 600 W of power, since what we expect in order to reduce the energy consumption of internal logistics that NVIDIA develops a new organization different from the one of GPC-TPC-SM that we have been seeing for more than ten years.
Hopper is expected by 2022
Let’s not forget that for the moment the multichip GPU for HPC under the code name Hopper is still promising for the future. Meanwhile AMD with its Instinct MI200 and Intel with its Ponte Vecchio have presented prototypes and one of them even a fully functional version. Nor can we forget how one of Bill Dally’s greatest obsessions at NVIDIA has always been the energy consumption of the operations that are executed in it, especially those related to data logistics.
The Hopper architecture is supposed to show up next year, but considering the timing between the V100 and the A100, an eventual H100 could take at least a little longer to show up. In any case, the use of monolithic chips has a limit on scaling and sooner or later NVIDIA will have to follow in the footsteps of its rivals by offering a multi-chip GPU.
The latest information speaks that Hopper could appear at the same time as Lovelace, simultaneous launches are not normal in this industry, so we could see it sometime in the first half of 2022 or already in the fourth quarter. NVIDIA typically runs two GTCs a year, so it is possible that they will feature Hopper in one of them.