We will wait for confirmation from NVIDIA, but the main characteristics of the next graphics cards GeForce suggest wise developments.
While NVIDIA maintains the mystery around the release “in 2024” of its next generation of graphics cards, there is no doubt that these GeForce RTX 50 series are very close to us anyway.
The question is rather knowing when exactly these Blackwell generation cards will arrive, and what the exact specifications of the associated GPUs will be. For this second point, it becomes clearer.
GB202: a monster of power!
Accustomed to leaks around NVIDIA GPUs, it is the informant kopite7kimi who, on X.com as always, has just made some juicy revelations detailing practically all the chips envisaged by the American firm.
There is no particular change regarding the most powerful GPU in the range, the GB202. In principle intended to equip the GeForce RTX 5090 card, this power monster will integrate 12 GPC (Graphics Processing Cluster) and 96 TPC (Texture Processing Clusters), for a more telling total of 192 flow multiprocessors (SM, streaming multiprocessors).
Kopite7kimi confirms the presence of 24,576 CUDA cores and, more importantly given the latest rumors, also confirms the presence of a 512-bit bus for video memory, which will be GDDR7. This last point is crucial as it marks progress on the current generation. The RTX 4090 makes do with a 384-bit memory bus.
Ada Lovelace optimizations?
The downgrade begins with the GB203, which we can imagine on a hypothetical GeForce RTX 5080. Its 7 GPC and 42 TPC imply the presence of 84 SM and a total of 10,752 CUDA cores. If the VRAM remains in GDDR7, it would only be a 256-bit bus.
With the GB205, we go to 5 GPC and 25 TPC, for a total of 50 SM and another 6,400 CUDA cores, but the VRAM is only on a 192-bit bus. The bus has been further lowered for the GB206 and GB207, which must make do with a 128-bit interface. The first GPU will have 3 GPC, 18 TPC, 36 SM and 4,608 CUDA cores, while the second has 2 GPC, 10 TPC, 20 SM and 2,560 CUDA cores.
Beyond just the values, the interest of this information is also to compare it to the data of the current generation, the GeForce RTX series 40. It is also interesting to note that apart from the GB202, the specifications are rather on the decline. Proof that Blackwell is above all an opportunity to optimize the Ada Lovelace architecture of our current cards?
2061.76€
9
- Maddening performance!
- DLSS 3 very convincing
- Rather discreet ventilation
Is it the arrival of Intel in this competitive market or the successive releases of AMD’s RDNA solutions that encourage NVIDIA to push the limits? No doubt a little of both, surely also that the company’s engineers have many good ideas to improve their architectures. Yet while CPUs are no longer making massive progress, GPUs seem to have some room to spare.
It seems clear that the Lovelace architecture has a bright future as the improvements observed are decisive, starting of course with DLSS 3. NVIDIA having once again opted for a proprietary solution, it will however need to ensure it. follow-up with developers, who will perhaps be tempted by calls from AMD or Intel with solutions that are certainly less advanced – significantly less even – but also much more open.
Lovelace appears to be the most powerful, and also the most efficient, architecture. While energy sobriety is on everyone’s lips, it bodes well for the other variations of Lovelace. A powerful monster, it is oversized for video game use only unless you plan to play in 8K. On the other hand, its power reserve, its quantity of memory and its software environment make it a remarkable semi-professional tool.
Is it the arrival of Intel in this competitive market or the successive releases of AMD’s RDNA solutions that encourage NVIDIA to push the limits? No doubt a little of both, surely also that the company’s engineers have many good ideas to improve their architectures. Yet while CPUs are no longer making massive progress, GPUs seem to have some room to spare.
It seems clear that the Lovelace architecture has a bright future as the improvements observed are decisive, starting of course with DLSS 3. NVIDIA having once again opted for a proprietary solution, it will however need to ensure it. follow-up with developers, who will perhaps be tempted by calls from AMD or Intel with solutions that are certainly less advanced – significantly less even – but also much more open.
Lovelace appears to be the most powerful, and also the most efficient, architecture. While energy sobriety is on everyone’s lips, it bodes well for the other variations of Lovelace. A powerful monster, it is oversized for video game use only unless you plan to play in 8K. On the other hand, its power reserve, its quantity of memory and its software environment make it a remarkable semi-professional tool.
Source : VideoCardz
9