Upcoming Nvidia RTX 5070 GPU rumoured to have fewer CUDA cores than the RTX 4070 Super

pcgamer.com:

Hot on the heels of details of the supposed RTX 5070 Ti GPU come details of the next rung down in the rumoured Nvidia RTX 50 family of graphics cards. We give you the RTX 5070, the plain vanilla version and what could be a worryingly small step forward if this new «leak» is anything to go by.

According to Benchlife (via Videocardz), we're looking at 6,400 CUDA cores for the RTX 5070. That's well down on the 8,960 rumoured for the RTX 5070 Ti. Perhaps even more pertinently, it's just 9% more CUDA cores than the existing RTX 4070 and, indeed, fewer CUDA cores than the RTX 4070 Super, which rocks 7,168.

Now, there is more to GPU performance than mere core count. There's clock speed, which you might expect to be up at least a little bit thanks to the transition from TSMC's N5 silicon to a revised N4 node for the Blackwell architecture that will underpin all GPUs in the RTX 50 lineup.

That said, N4 is not a wholly new node compared to N5, more of an enhancement. Nvidia would have needed to make the jump to N3 for that. So, it may not be realistic to anticipate a huge jump in frequencies.

That said, the new 5070 is rumoured to be rated at a board power of 250 W, which is a fair old step up on the 200 W of the 4070. So, we'll wait to see about those clock speeds and cross our fingers and toes that they're up a decent chunk.

Then there's architecture. It may turn out that the Blackwell in the RTX 5070 is much more efficient than the Ada Lovelace technology in the RTX 4070 and indeed 4070 Super. So, Blackwell may do more work per clock cycle. Features like the RT cores for ray tracing and the Tensor cores for AI stuff may get a particularly big boost. In other words, the RT 4070's CUDA cores and broader per-unit performance probably won't be perfectly comparable with those of the RTX 4070 and 4070 Super.

Even so, if all these rumours are correct, the gap between the RTX 5070 and 5070 Ti in terms of CUDA core count is 40%, whereas the same metric for the RTX 4070 and 4070 Ti