A more reliable insider with a well-deserved reputation, Kopite7kimi, spoke in detail about two NVIDIA “Ada Lovelace” video cards, GeForce RTX 4090 i GeForce RTX 4070.
He’s already revealed the alleged specs of other cards like the RTX 4090 Ti and RTX 4080, and here’s more information on the two cards mentioned above. The main difference is that the GeForce RTX 4090 will be based on an optimized graphics core. AD102while the RTX 4070 will use an optimized graphics coreAD104.
NVIDIA GeForce RTX 4090 it turns out that it will be based on the same GPU AD102, as is the “RTX 4090 Ti” variant, though rumored to be in a stripped-down configuration. Optimized GPU AD102 will use 144 shader units, which is 71% more than the existing GA102 GPU, and also contains 18432 CUDA cores. In addition to shader models, “Ada Lovelace” GPUs will also have an increased cache size. L2. Starting with the AD102 GPU, the flagship will be equipped with up to 96 MB cache memory in the second level, which is 16 times more than the GA102 with 6 MB second level cache.
The video card will also carry on board 24 GB memory GDDR6X clocked 21 Gbps with interface 384-bit tires and throughput more than 1.0 TB/s. Judging by the TGP parameter that appeared before, the card will have a reference power TBP 600 W.and it will increase in custom designs.
Looking at the GeForce RTX 3090 Ti for example, some models are out of scope 450 W TBP in the standard version, and 516 W. in non-standard. The same will happen with cards based on “Ada Lovelace”. And while this will require a lot of cooling, it seems that the board partners are well prepared. There will likely be more hybrid-style GPU coolers being used rather than traditional air coolers once the 600W TBP limit is exceeded. This means a 33% increase in TBP compared to the RTX 3090 Ti.
class cards NVIDIA GeForce RTX 4070will use GPUs AD104which will have 60 shader model blocks, in total 7680 CUDA cores i 48 MB L2 cache. Specifically, the RTX 4070 series will have 12 GB memory GDDR6working through 192-bit bus interfaceand TGP power 300 W. Given that these are GDDR6 chips, you get a memory speed of 16-18 Gbps, which should provide a throughput of about 400 GB / s. Most likely, this card is formed on the PG141-SKU341 printed circuit board.
Considering TDP 300 W., GeForce RTX 3070 Ti already has a reference TDP of 290W, so a negligible increase of just 10W is expected, but at the same time, performance is said to be on par with an RTX 3080 Ti-class graphics card that is rated at 350W. The 12GB memory also sounds very attractive as it is the same amount as the RTX 3080 Ti. Everything will depend on the price of this card, since the class RTX 3070 is one of the most popular among top-end graphics cards, and for users to get the performance of the RTX 3080 Ti for $1199, already at a price of about $500-600 would be a huge plus.
NVIDIA’s “Ada Lovelace” family of GPUs will make the leap between generations, like from “Maxwell” to “Pascal”. It is expected to be released in the second half of 2022, but at the same time, shipments and prices will be similar to the current cards, despite NVIDIA having a huge spend to acquire TSMC transistor wafers with 5nm (4N) process technology.