NVIDIA RTX 4090
With the same industrial design as the current generation, but with a bigger fan, NVIDIA showed us the first image of its RTX 4090. Promising to be between 2 and 4 times faster than the 3090 Ti, the NVIDIA graphics card most powerful to date. . The amount of memory it carries is 24GB GDDR6X at a speed of 21Gbps, so it has the same bandwidth as the RTX 3090 Ti, 1TB/s.
Regarding its price, in the United States will be 1599 dollarsin exchange in Europe we will have to pay 1959 $ to have at least one.
NVIDIA RTX 4080
At the same time, NVIDIA showed the RTX4080However, we know that despite the homonymous name, these are not two graphics cards with the same GPU in terms of configuration. Of course, both share the same architecture, but one has 12 GB of video memory, the other 16 GB. one of them uses GDDR6X at 22.5 Gbpsthe 16 GB one, but we don’t know if this is the case for the other model.
When it comes to pricing, we’re in another bucket of cold water, as the recommended retail price for the 16GB version is $1,199 in the US. While here in Europe the 12 GB will be available for $1,099 and the 16 GB will cost $1,469 As for its release date, it should be out in november. When it comes to performance, NVIDIA promises 2-4 times the performance of the RTX 3080 Tibut you didn’t specify which of the two models you are referring to exactly.
RTX 4000 Ada Lovelace Architecture Improvements
In addition, NVIDIA showed the following news regarding the architecture of its new family of graphics cards, some of them we must admit surprised us and we did not expect them, others, on the other hand, are an incremental improvement of what has already been seen before, but in the same way what they bring again has impressed us very pleasantly.
Incremental chip improvements
For the moment and in the form of a slide, they have given us information on the raw performance of the AD102 chip, the one that will ship the RTX 4090 and the most powerful to date under the new Ada Lovelace architecture.
Among the improvements we can highlight:
- 76 billion transistors.
- 90 TFLOPS of power in FP32 with up to twice the efficiency of the previous generation.
- 3rd Generation RT Corewith a computing power equivalent to 200 TFLOPS and with the ability to calculate twice as many intersections box-ray and triangle-ray than the previous generation.
- New Tensor Core, with 1400 TFLOPS oh 1.4 Power PetaFLOPS. Data support in FP8.
- New optical flow accelerator, of which they have not given any data yet, but it is unprecedented hardware so far in a GPU.
Reorganization of shader execution
under this name Reorganization of shader execution NVIDIA has announced that RTX 40s with the Ada Lovelace architecture are the first GPUs in history with out-of-order execution. A way of executing instructions that has existed in PC processors since the mid-1990s and shows how GPU cores have evolved so far in terms of complexity.
Nvidia has therefore taken a giant step by completely abandoning the classic Round-Robin method used since the first graphics cards with programmable shaders. In the realm of running code on a GPU in the form of shader programs, whether for graphics or computing, this is one of the biggest leaps ever in the history of graphics hardware. .
It also completely breaks comparisons when it comes to TFLOPS, since differences between a failed processor and one that is not under the same computing power are not ruled out at all.
DSLS 3
One of the improvements introduced by NVIDIA is the third generation of its automatic resolution download algorithm via deep learning. In the third version of which they claim to be able to quadruple the frame rate per second compared to rendering a scene in native resolution.
The key to this improvement over DLSS 3 are two: the Ooptical flow accelerator and the Multi-frame optical generator, which are added as elements to those already implemented in DLSS 2.0. The use of the optical flow accelerator and the fact that NVIDIA mentioned it as part of the circuitry of the RTX 4000 chips is confirmation that DLSS 3 is exclusive to the RTX 4000.
Ray Tracing Improvements
Faced with Ray Tracing according to NVIDIA, we have gone from 39 operations per pixel of the NVIDIA RTX 2080 Ti released in 2018 up to 635 operations per pixel equivalent of the RTX 4090 that NVIDIA has just presented. A 60x increase over the original.
Jen Hsen Huang made it clear that much of the performance gain comes from prediction via AI and DLSS 3, not just raw power. Thus, most pixels are predicted by the AI and not calculated.
Remix RTX
NVIDIA presented Remix RTXan application that allows capture old games across the Omniverse and allow us to recreate them again with new technologies such as Ray Tracing. When we saw this technology, we couldn’t believe how impressive it is. How many classic games will be remade and updated by creators and fans?
At the moment we have the first example with Portal with RTXa remaster of the classic Valve puzzle powered by Omniverse and rendered entirely via Ray Tracing and NVIDIA DLSS 3. It will be released in November for free for owners of the game on Steam.
Table of Contents
