The GeForce RTX 4090 will be available in less than a month, starting at $1599.
As expected, Jensen Huang, CEO of NVIDIA, formalized two GeForce RTX 4000 during its keynote from GTC 2022 : the GeForce RTX 4090 and RTX 4080. These cards carry an Ada Lovelace GPU engraved on TSMC’s N4 process. They benefit from 3rd generation RT cores and 4th generation Tensor cores.
In addition, the company announced the arrival of the DLSS 3.0 (Deep Learning Super Sampling); we will come back to this in another article.
From NV1 to the GeForce RTX 4090, here is the story of NVIDIA chipsets and graphics cards
GeForce RTX 4090 in October
The GeForce RTX 4090 is powered by an AD102-300 GPU. This one has good 75 billion transistors. The card has 16,384 CUDA cores; for comparison, GeForce RTX 3090 at 10,496, the GeForce RTX 3090 Ti, 10,752. With a Boost frequency of 2520 MHz, the card achieves a theoretical single precision power of 82.6 TFLOPS, more than 2.3X that of the RTX 3090 (35.58 TFLOPS). On the VRAM side, there are 24 GB of GDDR6X, for a memory bandwidth of just over 1 TB/s. This monster will be available from October 12, starting at 1599 dollars.
GeForce RTX 4080 in November
The GeForce RTX 4080 16 GB has an AD103-300 GPU. This reference has significantly fewer CUDA cores than its big sister: 9728. It has 16 GB of GDDR6X for a bandwidth of 720 GB/s. The GeForce RTX 4080 will arrive later, in November, at a recommended price of $1199.
NVIDIA will also offer a 12GB version starting at $899. This card is not only amputated at the level of its VRAM, but would also offer fewer CUDA cores (7680) via an AD104-400 GPU; in addition, its memory speed would be lowered to 21 Gbit/s. However, these data require confirmation.
Finally, no GeForce RTX 4070 – yet.
Graphic card | GeForce RTX 4090 | GeForce RTX 4080 16 Go | GeForce RTX 4080 12 Go |
SKU | PG139-SKU330 | PG139-SKU360 | PG141-SKU331 |
Architecture | Yes (TSMC 4N) | Yes (TSMC 4N) | Yes (TSMC 4N) |
GPU | AD102-300 | AD103-300 | AD104-400 |
Stream Multiprocessors | 128 | 76 | 60 |
Hearts CUDA | 16384 | 9728 | 7680 |
Frequency boost | 2520 MHz | 2505 MHz | 2610 MHz |
VRAM | 24 Go G6X | 16 Go G6X | 12 Go G6X |
Memory bus | 384-bit | 256-bit | 192-bit |
Memory speed | 21 Gbit/s | 22,5 Gbit/s | 21 Git/s |
Memory bandwidth | 1008 Go/s | 720 Go/s | 504 Go/s |
Default TGP | 450W | 320W | 285W |
Max TGP | 660W | 516W | 366W |
MSRP | 1599 dollars | 1199 dollars | 899 dollars |