Nvidia and AMD are currently working on their next generations of GeForce RTX and Radeon RX. New rumors appear. They address a variety of issues, including that of video memory.
The next family of GeForce RTX graphics cards won’t necessarily be called the GeForce RTX 40 series. This name has now been “temporized” by various indiscretions. Nvidia will likely adopt the GeForce RTX 50 series name for its generation Ada Lovelace† kopite7kim list these solutions as **90/**80 etc, in case they are not called RTX 40 but RTX 50.
GeForce RTX Ada LoveLace
The new speculation around video memory does not call for 24 Gbps but 21 Gbps. We would suddenly have a profile identical to that proposed by the new one GeForce RTX 3090 Ti† The thermal envelope should reach 600 Watts. This figure is plausible regarding the PCIe 5.0 connector mounted on the current 3090 Ti.
GeForce RTX “Ada Lovelace”.
| GPUs | GA102 | AD102 |
| architecture | Ampere | Ada Lovelace |
| Engraving | 8nm | 5nm |
| Graphics Processing Clusters (GPC) | 7 | 12 |
| Texture Processing Clusters (TPC) | 42 | 72 |
| Streaming Multiprocessors (SM) | 84 | 144 |
| CUDA cores | 10.752 | 18,432 |
| L2 cache | 6MB | 96MB |
| Power TFLOPs | 37.6 | 90? |
| Memory type: | GDDR6X | GDDR6X |
| Memory bus | 384 bit | 384 bit |
| Memory capacity | 24GB (3090) | 24GB (4090 or 5090?) |
| TGP | 350W | 600W |
| Publication date | September 20 | 2H 2022? |
Based on rumours
the GPU Ada Lovelace AD102 would consist of 144 SMs, a 71% increase over the current GA102 GPU! At the heart of its mechanics, no less than 18,432 CUDA cores would be in action, as well as even a 384-bit bus, which is enough to guarantee 1004 Gb/s bandwidth.
GeForce RTX 5090 or RTX 4090, a monster equipped with an AD102 GPU and 24 GB of memory
Radeon RX RDNA 3 (Radeon RX 7950 XT?)
On AMD’s side it mumble that the next Navi 31, the flagship of the RDNA 3 family (Radeon RX 7950 XT?) will be equipped with 7 chiplets. Not everything is very clear on this subject. Normally AMD would use an MCM contraction approach of Multi Chip Module with 2 GCDs (5 nm) and 4 MCDs (6 nm), all accompanied by an IOD.
This hypothetical Radeon RX 7950 XT, its Navi 31 GPU, should offer 15,360 stream processors operating at 2.5 GHz. We don’t know whether this frequency represents the maximum possible value or the base value. Meanwhile, this is an 11% improvement over the Radeon RX 6900 XT, enough to ensure, with other changes, a theoretical power of 38.4 TFLOPs in FP32, a nice +67% compared to the Radeon.RX6900XT. The card should be accompanied by an impressive TDP of 500 Watts, another 67% increase over its predecessor.
In terms of equipment, it cannot be ruled out that the working memory offers 32 GB, which is twice as much as the current AMD flagship. The operation of this GDDR6 at 21 Gbps would be ensured by a 256-bit bus. Likewise, the Infinity Cache should also double (512MB).
Here’s a signed first rumor-based report VideoCardz†
| GeForce RTX 4090 | Radeon RX 7950XT | |
| Fineness of engraving |
TSMC N5 |
TSMC N5/N6 |
| architecture | Ada Lovelace | RDNA3 |
| Design | monolithic | Multi-chip module (MCM) |
| dies | 1 | 2 GCDs + 4 MCDs + 1 IOD |
| MEGA clusters | 12 GPC (Graphics Processing Cluster) | 2×3 Shader motors |
| SUPER Clusters | 72 TPCs (Texture Processing Clusters) | 2×30 WGP (RDNA) |
| GPU clusters | 144 text messages | † |
| FP32 cores | 18,432 CUDA cores | 15,360 Stream Processors |
| GPU Frequency | ~2.2GHz | ~2.5GHz |
| RAM | 24GB GDDR6X | † GBGDDR6 |
| Memory frequency and memory bus | 21 Gbps – 384-bit | † Gbit/s – 256-bit |
| Hidden | 96MB L2 cache | 256 or 512 MB Infinity Cache |
| Consumption | 600W (TDP) | 500W (TBP) |
| Publication date | Q3/Q4 2022 | Q3/Q4 2022 |
| Performance | 2x GA102 |
2.5x NAVI 21 |
