NVIDIA GeForce RTX 4070 Specs Allegedly In Limbo With Two Designs Under Consideration
Back in June, kopite7kimi updated his leaked specifications for what are likely to be the top three GeForces in the RTX 4000 range, at least at launch: the RTX 4070, RTX 4080, and RTX 4090. At that time, he predicted that the RTX 4070 would use an AD104-275 GPU with 7168 shaders connected to 10 GB of 18 Gbps GDDR6 memory over a 160-bit bus. He predicted that such a card would draw around 300 watts.
That's similar to one of the configurations he's predicting now, except that he's bumping the memory spec to 21 Gbps GDDR6X and dropping the total card power limit to 250 watts. This would still be quite a powerful GPU even if we were talking about an Ampere card, but we have little trouble believing that such a card could compete with the RTX 3080 as promised in July with the Ada architecture's improvements.
The other proposed configuration is beefier, making use of a full-fat AD104 GPU die with 7680 shaders hooked up to 12 GB of GDDR6X memory across a 192-bit bus running at 21 Gbps. That's a substantial upgrade, less in terms of GPU compute and more in terms of memory bandwidth: from 420 GB/sec of the lower spec up to 504 GB/sec on the higher spec. We have to note, though, that both numbers are behind the GeForce RTX 3070 Ti's memory bandwidth.
Updated chart. Note that 21 Gbps memory is faster than everything out now besides the RTX 3090 Ti.
While we'd certainly prefer to see at least 12GB of video RAM on an x70 card—especially after the awkwardness of having already seen it on an x60 card—we find the idea of NVIDIA using a fully-enabled AD104 GPU on the RTX 4070 to be a little dubious. That means that if the company wanted to introduce a "Ti" or "SUPER" model later, it would have to use cut-down AD103 dice, which are more expensive. That's not unheard-of, but it is uncommon.
We also still find the rumored memory bandwidth specs on the Ada parts as a whole to be a little anemic. It's possible that the massively increased last-level cache of the Ada architecture reduces the GPUs' reliance on having massive main memory bandwidth. For gaming, at least, it's a possibility, although it won't help for compute workloads that are sensitive to memory performance. Of course, NVIDIA would probably rather sell you a Quadro for those tasks, anyway.
