NVIDIA GeForce RTX 4000: information and rumors about Lovelace and Hopper
Starting with the rumors, here’s what we know about the next family of NVIDIA video cards: the GeForce RTX 4000.
For an abundant week there has been strong talk of the next generation of NVIDIA video cards, the future GeForce RTX 4000 which will replace the 3000 family based on the Ampere architecture.
We absolutely realize how absurd it is to talk about the 4000 or 3000 Super Series already, when even today it is practically impossible to bring home a 3070, a 3080 or a brand new 3060 Ti. But this is how the technology market goes: whoever stops is lost and it is essential to constantly look forward.
Based on a whole series of leaks, rumors, rumors and rumors that have started to bounce since mid-December, it seems that NVIDIA has substantially changed its roadmap, on the one hand accelerating with the aim of hindering the inexorable advance of AMD and on the other postponing what already promises to be a radical change in the production system and the design of its GPUs.
Today we will therefore take stock of the situation on Lovelace, the code name of the next architecture that will be the basis of the GeForce RTX 4000. We will make some assumptions about when it should arrive on the market and we will tell you about the specifications that, we already tell you, seem science fiction.
However, we will not fail to tell you some details also of Hopper, the subsequent architecture, the one that should represent a radical change in NVIDIA’s GPUs and which, apparently, will take longer than expected.
For more than a decade now, NVIDIA has chosen to name its architectures after famous and revolutionary scientists, mathematicians and inventors of the past, not necessarily related to computer science but still related to some, important, technological progress. And so, after a very long period of male exponents, we recall among the most recent Pascal, Turing and Ampere, respectively linked to the GeForce GTX 1000 and the RTX 2000 and 3000, waiting for us in the future there will be 2 architectures dedicated to female scientists: Ada Lovelace and Grace Hopper.
Lovelace is considered by many scholars to be the first, true fanatic of computerized computing, even if in her time the computer was not even hypothesized. He was the first person to theorize that Charles Babbage’s analytical engine, a sort of computer prototype never actually built, could be used to go beyond pure and simple mathematical calculations and, in this regard, he wrote what is considered the first true algorithm, effectively becoming humanity’s first software programmer. And the beauty is that it became one half a century before Alan Turing built the first programmable computer during World War II.
There is actually a bit of indecision about the name of this architecture. According to a handful of stolen internal documents, NVIDIA refers to the new GPU with the code name Ada, while in the old official presentations that already presented hypotheses of names of future architectures, there was talk of Lovelace. Eventually, NVIDIA will likely focus on the shortest and simplest name: Ada, leaving aside the scientist’s surname.
The first, very important breaking point with the past is in productive process. The current 3000 Ampere family GPUs are manufactured in Samsung foundries with an 8 nanometer production process which seems to have been one of the main reasons for the delays in the distribution of video cards on the market, in addition to Covid-19, and the main cause of the scarcity of supplies for a production yield much lower than expected.
It therefore seems that NVIDIA has decided to take the step much longer than the leg, leaving behind not only the 8 nanometers of Samsung, but also the 7 nanometers of TSMC, the other well-known foundry on the market, used by AMD for its most recent Radeon video cards and for Ryzen processors.
The goal of the Santa Clara house is to aim directly at the 5 nanometers, even if at the moment no one seems to know NVIDIA’s partner, that is, if it will continue to rely on Samsung, or if it will point to TSMC by opting for the exclusivity of the production process so as to take advantage of a manufacturer that has proved to be more reliable and faster and also putting the spanner in the works for AMD that would find itself without its historic ally for future RDNA 3-based video cards.
As if this were not enough, it seems that NVIDIA has placed an order for the production of a large quantity of GPUs with the production process at 6 nanometers. A sort of middle ground that allows you to take back the scepter of the video card with the most advanced production process on the market. The problem is that even in this case it is not clear whether it is a sort of revision of Ampere, perhaps the 3000 Super series along the lines of what was done with the 2000 family, or whether it is a completely new chip, perhaps relative. to other market segments.
What is now taken for granted is that Lovelace will use the 5 nanometer process and will have specifications simply phantasmagoric.
The main chip in fact, what is believed to be called AD102, should have an internal design based on a 12×6 structure as opposed to the 7×6 design that we find instead on Ampere’s GA102, the one at the base of the RTX 3080 and 3090. Which would translate into 12 Graphic Processing Clusters and 72 Texture Processing Clusters. The latter would allow to have 144 Streaming Multiprocessors for a total of 18432 CUDA Core, just under double the CUDA Cores of the top-of-the-range Ampere chip which stops at 10752. Assuming a working frequency of 1.8 GHz, we should reach 66 TeraFLOPS maximums in 32-bit floating point calculations, a leap forward that nearly double the 35-odd TeraFLOPS achieved by the 3090.
These astounding numbers should be allowed precisely by the substantial leap forward as regards the production process. Going from 8 to 5 nanometers means being able to increase the size of the chip, be able to increase the number of transistors per square millimeter, have lower energy consumption and less heat production. And therefore to be able to work on a consistent increase in the clusters that make up the GPU and maybe even retouch the working frequencies upwards.
And all this without taking into consideration the second-level cache which, even in this case paying attention to the rumors, should be substantially increased after being blocked for 2 generations in a row. All beautifully shining? Apparently yes, but clearly a huge question remains.
When will arrive on the market the new video cards? And here the toto betting literally opens since there are so many variables at stake and, above all, we are still in the midst of a pandemic that has literally upset the roadmaps of all the major technology producers.
Today we still have to deal with the very limited availability of video cards of the 3000 family. And it is very likely that some models are still missing: a 3050 for the very low end of the market, the 3070 Ti and the 3080 Ti or, even a refresh after a year or less, as done with the 2000 family which could return to using the Super suffix and could further increase the ratio between price and performance perhaps through the boost of frequencies and the increase of RAM.
And we also insert in the cauldron the speech of the mobile version of the 3000 which, to this day, is missing and which could become the protagonist of the CES in Las Vegas in January.
If we want to make some assumptions, we would like to say that for the whole of 2021 NVIDIA will continue to squeeze the Ampere architecture by introducing on the market from now to spring the mobile versions and the latest models remained excluded such as the Ti versions of the 3070 and 3080. Then, on the at the end of the year it will be the turn of the lineup refresh with the Super models that could also compensate for the reduced availability of all these months.
And only in the 2022, at the turn of autumn, we will finally be able to see, in our opinion, the 4000 models.
Hopper: the architecture of the future
But as we told you at the beginning of the video, we are not here only to talk to you about Lovelace since the changes to the NVIDIA roadmap also concern the subsequent architecture, the one that should completely revolutionize the structure of the GPUs of the Santa Clara house. and above all it should allow new progress in the decades to come.
We are talking about Hopper, the technology that seems to have been moved a few years forward, precisely to allow Lovelace to advance. Here the details are very scarce and focus exclusively on the extent of this Revolution which, touching the internal design of the die, should result in a radical change in the production of video cards. A change that AMD has actually already successfully embraced with its Ryzen processors and which should be an integral part of RDNA 3.
Here we risk going into a lot of detail. For now, just know that Hopper should involve switching from monolithic design current to that MCM, multi-chip.
To put it as simply as possible, but not necessarily in the most correct way, imagine that today the chips at the base of NVIDIA video cards are designed as a single, closed system: a single die of ever-increasing size that contains all the architecture. and all the gears, if you pass me the term, inside it. When the chip is printed on silicon wafers, as its size increases, the risk that it will not print correctly and therefore be defective increases.
And, among other things, the more the die grows, a typical thing of the transition between generations and architectures, the less you can print on the wafer. And therefore you risk throwing away more and more.
The MCM design instead provides that the chip is broken into many much smaller parts that can then be mounted in series, as for the processors, or in parallel, as for the video cards, inside the final SoC. By doing so on the same wafer you can print many more components, since they are smaller, you risk less regardless of the yield of the production system and above all you can proceed virtually indefinitely with the enlargement of the die. In any case, what matters is to assemble these small parts in gradually increasing numbers.
Here, Hopper should really represent this very important generational leap that would allow NVIDIA to look to the future with new and increasingly consistent and larger architectures, without worrying too much about production yield.
Ah, just for the sake of completeness, Grace Hopper was another woman who pioneered information technology and computers. She was one of the first programmers of the Harvard Mark 1, one of the very first electromechanical digital computers in history, built during the Second World War and was one of the main theorists of the concept of an architecture independent programming language that then led to the realization of COBOL , a language still used today.
Among other things, let’s never forget that NVIDIA has owned Arm for a few months now and, we are convinced, sooner or later they will start producing CPUs. Who knows that a unified architecture with processor and GPU on the same SoC may not benefit from an MCM design. But that’s another story and maybe we’ll talk about it in another article.
For now we stay focused on the upcoming GeForce RTX 4000. Are you curious to find out how strong these video cards will go? Or do you think technology is progressing at an excessive pace? And can’t it be exploited properly?
Let us know below in the comments.