Nvidia RTX 4090 Ti with 18,172 CUDA cores shelved due to lack of rivalry from AMD, Blackwell to feature 512-bit memory bus

So far, it looks like Nvidia has completed its RTX 40 series lineup of GPUs for desktop and mobile with the latest launches being that of the RTX 4060 and the two RTX 4060 Ti cards. The flagship AD102-based RTX 4090 features 16,384 CUDA cores against the 18,432 CUDA cores that is possible with the full AD102 silicon. 

This meant that an "RTX 4090 Ti" was in the offing sometime later this year, similar to what we had seen with the RTX 3090 and the RTX 3090 Ti. However, it looks like Nvidia might not offer any RTX 4090 successor when the card celebrates its first anniversary this October.

Well-known and reliable Nvidia leaker @kopite7kimi has tweeted that there will not be an RTX 4090 Ti, but Nvidia might introduce new RTX 4070 and RTX 4060-class cards based on AD103 and AD106 refreshes, respectively.

It is not fully clear why Nvidia has decided to forego an RTX 4090 successor in the Ada Lovelace generation. The likely reason is that there is no competition to the RTX 4090 from AMD. We did hear rumors that Nvidia was thinking of a contingency in the form of a GH202 Hopper in case AMD went all guns blazing with RDNA 3, but AMD seemed content playing second fiddle this time.

The closest one can get to experiencing complete AD102 is with the RTX 6000 Ada professional GPU that boasts of 18,176 CUDA cores. We did come across a prototype RTX Titan Ada with a PG137-SKU0 board, AD102-450 GPU featuring 18,176 CUDA cores, and 48 GB GDDR6X VRAM, but Nvidia looks to have shelved this monstrous 4-slot card. There were even reports of this card melting power supplies and tripping circuit breakers thanks to its purported 800 W TGP requirement.

That being said, it is becoming clearer that we will not be seeing a new Ada flagship anytime soon. The interest is gradually shifting towards Ada's successor, touted to be the RTX 50 Blackwell. From what we know so far, Nvidia is planning a 2.6x performance bump from Ada.

Blackwell is still in its infancy, but word on the street is that Nvidia engineers are mulling to increase streaming multiprocessors or even a multi-chip design in the event such a performance boost is not attainable. Now, we are also getting to know that Blackwell will make use of a 512-bit memory interface.

Whether such a memory bus will be used in conjunction with GDDR7 memory and in which SKUs it would be enabled will only be known once the RTX 50 lineup becomes clearer in 2025.

Buy the Aorus GeForce RTX 4090 Master GPU on Amazon