The leaks released on the next series of GeForce RTX 3000 video cards allow to outline the profile of the Ampere GPUs.
Summer has always been the time of year when leaks and rumors about the world of technology and video games embrace each other, in an infinite spiral of web chatter. The astral conjunction between the presentation of the new consoles by Microsoft and Sony and the arrival of the next generation of NVIDIA and AMD GPUs completes the scenario of this strange and very interesting summer 2020.
If we know something about the RDNA2 architecture (still too little) thanks to Xbox Series X and PlayStation 5, for consumer-grade Ampere GPUs we have to rely on the words of the various insiders and specialized technicians. So let’s try to take stock of the situation on the new proposal of NVIDIA and on the so-called graphics cards of the GeForce RTX 3000 series. Between dreams, hopes and solid realities.
The kitchen of Jensen Huang
During the latest NVIDIA showcase held in May, CEO of the company Jensen Huang presented the first Ampere GPU to the world for scientific and industrial purposes. The A100 GPU is a 7-nanometer chip that contains 54 billion transistors in an 826 mm2 die (to make a comparison with the current generation of consumer cards, the GeForce RTX 2080 Ti contains 19 billion transistors). According to NVIDIA’s statements, A100 is capable of elaborate about 19.5 TeraFLOPs of floating point calculations thanks also to a 40 GB HBM2 memory with a bandwidth of 1.5 TB / s. The product in question is obviously intended for use in servers and scientific calculators and is in fact contained in the new supercomputer DGX A100.
The company said that in performance, a single A100 GPU can replace an entire rack of servers based on current NVIDIA Volta. However, what interests gamers are the innovations brought by this new technology, which bode well for the consumer: the chip introduces support for TF32 and the PCIe 4.0 standard, in addition to a considerable optimization of the AI, a factor that in the future it could translate into better management of the DLSS and the much coveted Ray Tracing.
Let’s give the numbers
Over the past few months various insiders and leakers have provided more or less reliable information on the technical characteristics of the Ampere GPUs and on the possible performance of the new NVIDIA RTX 3000 graphics cards.
Thanks to the analyzes carried out by the technicians and the specialized press, we can outline the profile of the models that came to light with a certain precision and a rather low margin of error. For convenience we will summarize the data schematically:
NVIDIA GeForce RTX 3080 Ti / RTX 3090: GPU Ampere GA102-300, 5248 CUDA Core and 82 SMs, 12 GB DDR6X, 384 bit memory bus and 1 TB / s passband
NVIDIA GeForce RTX 3080: GPU Ampere GA102-200, 4352 CUDA Core and 68 SMs, 10 GB DDR6X, 320 bit memory bus and a passband of 760 GB / s
NVIDIA GeForce RTX 3070 Ti: GPU Ampere GA104-400, 3072 CUDA Core and 48 SMs, 8 GB DDR6X, 256 bit memory bus and a 512 GB / s bandwidth
NVIDIA GeForce RTX 3070: GPU Ampere GA104-300, 2944 CUDA Core and 46 SMs, 8 GB DDR6, 256 bit memory bus and a 512 GB / s passband
The appeal is still missing two models, the GA106-400 and the GA106-300, which they could represent respectively NVIDIA GeForce RTX 3060 Ti is RTX 3060, of which for the moment the technical details are unknown. The GA102-400 model, known as NVIDIA Titan RTX second generation , with 5376 Cuda cores and 84 SMs, 24 GB of DDR6X memory, 384 bit bus and 816 GB / s bandwidth. As the insider KatCorgi pointed out, if the numbers were confirmed, in the case of the RTX 3080 model we would be faced with an increase in performance compared to an RTX 2080 in the order of 20% (which could go up to 60% with the RTX 3090). In general, therefore, a decisive step forward from the point of view of mere specifications, a goal achieved thanks also to the new 7 nanometer production process adopted by NVIDIA for the next generation of GPUs.
For weeks now, the images of the alleged form factor of the new RTX 30 have been circulating, with a huge heatsink block placed centrally and a pair of axial fans that according to Igor’s Lab colleagues could even be among the most expensive ever produced.
The latest rumors have also highlighted the presence of a new 12-pin connector on all reference models (a factor that could affect compatibility with some power supplies).
The revolution of functionality
Beyond the technical characteristics and theoretical power, what has always distinguished the graphics cards of the GeForce series is the innumerable supply of features on which the success of NVIDIA has pivoted and which have often been synonymous with innovation. With the next hardware, the Californian company intends to take another step. The Ampere GPUs and the GeForce RTX 3000 will in fact introduce the DLSS 3.0, evolution of the current artificial intelligence-based rendering algorithm that accompanies the RTX 2000. According to the latest rumors, the DLSS 3.0 will work with any game with TAA simply passing through a new one Basic Game Ready Driver.
Developers will have to make specific programming for each game but the amount of work required should be more than halved compared to the current generation. In performance terms, the new algorithm should allow significant improvements even if the real effects will be different from game to game.
If the rumors were confirmed the contribution of DLSS 3.0 could turn into a performance boost with active Ray Tracing about 3 or 4 times higher than seen with DLSS 2.0 and with current NVIDIA cards.
Among the innovations implemented with the RTX 3000 that emerged thanks to the leaks, the Cross coprocessor which should be equipped on the flagship products of the new line, namely the GeForce RTX 3090 and 3080. The function of this chip is not yet clear but persistent rumors speak of a unit designed to support and push the performance of Ray Tracing, lightening the Central GPU from that task.
Still, NVIDIA could introduce a feature called NVCache, similar to AMD’s HBCC. This feature should allow even faster loading times from RAM and SSD, allowing dynamic use of the band from SSD, VRAM and DDR4 at the same time. Finally, the Tensor Memory Compression, which through the use of the Tensor Cores can optimize the use of VRAM by lightening it by 20 – 40%.
In short, the hopes and expectations for the new generation of NVIDIA graphics cards are supported by increasingly plausible data. Never before has such a wealth of information emerged on this occasion. The last word, however, is up to NVIDIA. More and more insistent voices want the release of the GeForce RTX 3000 in September, in time to welcome highly anticipated titles like Cyberpunk 2077 and Halo Infinite with open arms. All that remains is to wait anxiously.