Kopite7kimi, a well-known leaker on Twitter, has published what may be the first specifications of Nvidia's upcoming GeForce RTX 40-series graphics cards family. Since we are several months away from Nvidia's launch of its codenamed Ada Lovelace architecture, it's possible there will be changes, but given the leaker's track record, the specs are at least worth considering.
Nvidia's initial GeForce RTX 40-series family will consist of three graphics cards: the GeForce RTX 4090, based on the AD102-300 graphics processor paired with 24GB of GDDR6X memory; the GeForce RTX 4080, powered by the AD103-300 GPU mated with 16GB of GDDR6/GDDR6X memory; and the GeForce RTX 4070 featuring the AD104-275 chip equipped with 10GB of GDDR6.
The number of CUDA cores featured by Nvidia's Ada Lovelace GPUs show that the new graphics boards will likely have a considerable performance increase when compared to Nvidia's existing GeForce RTX 30-series, which are among best graphics cards around.
Rumored Nvidia GeForce RTX 40-Series Specifications
Row 0 - Cell 0 | GPU | FP32 CUDA Cores | Memory Configuration | TBP |
GeForce RTX 4090 | AD102-300 | 16384 | 24GB 384-bit 21GT/s GDDR6X | 450W |
GeForce RTX 4080 | AD103-300 | 10240 | 16GB 256-bit 18GT/s GDDR6? | 420W |
GeForce RTX 4070 | AD104-275 | 7168 | 10GB 160-bit 18GT/s GDDR5 | 300W |
All these initial GeForce RTX 40-series graphics boards will allegedly be quite power hungry with the top-of-the-range product featuring a thermal board power (TBP) of 450W and other rated for 420W and 300W, respectively. When it comes to power consumption, the information corroborates with earlier rumors indicating a major increase of TBP for Ada Lovelace family.
While power consumption of Nvidia's next-generation GeForce RTX 40-series products will increase compared to current-generation graphics cards, not everything is clear about pricing of the new boards. Kopite7kimi tells us not to expect GeForce RTX 4000 boards to feature lower MSRPs. Price is a particularly murky topic, keeping in mind higher production costs on TSMC's N5 process technology (or rather customized 4N) when compared to Samsung's 8LPP.
It is noteworthy that this time around Nvidia will use three different graphics processors for three initial cards. This might be a result of decent initial yields of Nvidia's Ada Lovelace GPUs at TSMC, or perhaps due to some changes in the company's internal chip design schedules. Anyhow, the gaps between the number of CUDA cores featured by the AD102, AD103 and AD104 GPUs might indicate that Nvidia will be very flexible with configurations for refresh (Ti-branded or Super-branded products).
Of course, Nvidia typically does not comment on specifications of unreleased products, so take the information with a grain of salt.