Nvidia's forthcoming GeForce RTX 40-series (Ada Lovelace) graphics cards continue to make the rounds in the rumor mill. Hardware tipster kopite7kimi (opens in new tab), who has an impeccable history with Ampere leaks, has revealed the potential memory configurations for the GeForce RTX 4090 and RTX 4070.
The GeForce RTX 4090 reportedly rocks the AD102 die and is likely the flagship silicon for Nvidia's next-generation lineup. The graphics card may arrive with 24GB of GDDR6X memory at 21 Gbps. That's the same recipe Nvidia used for the GeForce RTX 3090 Ti. Therefore, Nvidia may not push the memory limit for this generation, seemingly maxing out at 24GB as the chipmaker did on Ampere.
Assuming that the GeForce RTX 4090 retains the 384-bit memory interface as the GeForce RTX 3090 Ti, the graphics card would deliver a memory bandwidth up to 1 TBps. Thus the GeForce RTX 4090 only offers 7.7% higher memory bandwidth than its predecessor, the GeForce RTX 3090.
In addition to the memory upgrade, the GeForce RTX 4090 would also sport more CUDA cores. However, the exact amount is unknown for now. Our estimation is between 17,000 to 18,500. Logically, Nvidia would need to raise the power limit on the GeForce RTX 4090 to accommodate the extra CUDA cores. Early diagrams of the GeForce RTX 4090's feasible PCB show that the graphics card could pull up to 600W of power.
Nvidia GeForce RTX 40-Series Specifications
|Header Cell - Column 0||GeForce RTX 4090*||GeForce RTX 3090 Ti||GeForce RTX 3090||GeForce RTX 4080*||GeForce RTX 3080 12GB||GeForce RTX 3080||GeForce RTX 4070*||GeForce RTX 3070|
|VRAM Speed (Gbps)||21||21||19.5||?||19||19||18||14|
*Specifications are unconfirmed.
The GeForce RTX 4080, on the other hand, might tap into the AD103 silicon. The hardware leaker (opens in new tab) previously stated that it would wield 16GB of GDDR6X memory. That's a 60% more memory than the vanilla GeForce RTX 3080 and a 33.3% upgrade over the subsequent GeForce RTX 3080 12GB. Again, we're unsure of the GeForce RTX 4080's CUDA core count, but kopite7kimi believes that AD103 has a similar TGP (total graphics power) as GA102. The maximum limit on GA102 is 450W, meaning Nvidia may have to keep the GeForce RTX 4080's CUDA core count to around 10,000 to remain within GA102's TGP.
As for the GeForce RTX 4070, Nvidia could use the AD104 silicon for the graphics card. The number of CUDA cores may hover over the 7,500 mark, give or take. According to kopite7kimi, the GeForce RTX 4070 will arrive with 12GB of 18 Gbps GDDR6 memory. So it won't just have 50% higher memory capacity than the GeForce RTX 3070, but the GDDR6 chips are also faster.
The new GeForce RTX 40-series graphics cards will probably hit the market before the year is over. Assuming that Nvidia continues with the same cadence as Ampere, we could see GeForce RTX 4090 and RTX 4080 in September, with the slower models arriving months later.