New NVIDIA GeForce RTX 5000 graphics cards from the Blackwell family are expected to be released in 2025. For a long time, there have been sporadic leaks about VRAM memory, the lithography used, the planned chip markings and bits of information about the architecture of integrated circuits. Now, however, the first leaks about the specifications of the top GeForce RTX 5090 system have come to light. Without further ado, let’s take a look at the upcoming performance king.
NVIDIA GeForce RTX 5090 will probably get GDDR7 VRAM and a GB202 chip, which will have as much as 50% more cores and 78% more L2 cache than the AD102 GPU in GeForce RTX 4090.
NVIDIA Blackwell – there may be major design changes in the next generation of graphics chips
Nvidia GeForce RTX 5090 will probably be based on a full or slightly trimmed GB202 core, which will work with GDDR7 memory. According to user Panzerlied from Chiphell, the new top GPU will feature an increase in cores by as much as 50% compared to the AD102 system (GeForce RTX 4090). Specifically, we are talking about an increase from 16,384 CUDA cores to approximately 24,000 CUDA cores. We should also get more RT cores, as it will increase from 128 to 192. This may herald quite a large performance increase in standard rasterization and faster processing of ray tracing calculations in video games. This is certainly good news for those who complained about a slight increase in rasterization performance on the occasion of the premiere of Ada Lovelace family systems.
| NVIDIA GeForce RTX 5090 | NVIDIA GeForce RTX 4090 | |
| Architecture | Blackwell | Ada Lovelace |
| Graphic layout | GB202 | AD102 |
| Lithography | TSMC N3 (3nm) | TSMC N5 (5nm) |
| Core size | No data | 608 mm² |
| Transistors | No data | 76 billion |
| SP units | ~24000 | 16384 |
| TMU units | No data | 512 |
| ROP units | No data | 192 |
| RT units | 192 | 128 |
| Tensor units | No data | 512 |
| Base clock | No data | 2235MHz |
| Boost clock | ~2900MHz | 2520MHz |
| Memory clocking | 24000MHz | 21000MHz |
| Amount of VRAM | 24GB or 32GB | 24GB |
| VRAM type | GDDR7 | GDDR6X |
| VRAM memory bus | 384-bit or 512-bit | 384-bit |
| The amount of L2 cache | 128 MB | 72MB |
| Capacity | 1536 GB/s or 2048 GB/s | 1000 GB/s |
| Interface | No data | PCI-E 4.0 x16 |
| TGP coefficient | No data | 450 W |
NVIDIA GeForce RTX 5000 – Samsung is already sending the first samples of GDDR7 memory for graphics cards to the company
The boost clock of the GB202 chip is expected to reach 2900 MHz, although it is said that the chip should be able to break the 3 GHz barrier. This is supposed to be a 15% clock-to-clock increase with the older graphics system. Additionally, the second level cache will be expanded, we expect the capacity to increase from 72 MB to 128 MB. Certainly, additional low-level memory will speed up communication with the GDDR7 VRAM, which is expected to reach clock speeds of 24,000 MHz. The memory subsystem will probably have a 384-bit or 512-bit data bus, which will translate into transfers of 1536 GB/s or 2048 GB/s. GeForce RTX 5090 will also receive 24 GB or 32 GB of VRAM.
NVIDIA GeForce RTX 5000 – new details about the Blackwell graphics chips being prepared
All these changes are expected to translate into 1.7 times greater overall performance than the GeFroce RTX 4090 graphics card. Unfortunately, we do not yet know the remaining parameters of the upcoming NVIDIA flagship, but remember that we are currently at least a year away from the premiere of Blackwell family systems and a lot may happen. change yet. Therefore, we recommend that you approach the above specification table with caution, although the upcoming GeFroce RTX 5090 undoubtedly looks quite impressive and heralds large increases in performance.
Source: VideoCardz, Chiphell (@Panzerlied)
Tags: NVIDIA GeForce RTX learned specifications graphics card leaks announce large increase performance
