Nvidia RTX 4070 10GB prototype appears on GPU-Z — the configuration provided more shader cores but less memory and bandwidth

Did you know that the RTX 4070 was originally designed with just 10GB of VRAM in mind? So go the rumors, and Harukaze5719 on X (Twitter) discovered one of these ultra-rare RTX 4070 prototypes featuring a 10GB configuration, rocking the AD104-275 die and with a 160-bit memory interface. The RTX 4070 that actually launched has 12GB of GDDR6X memory on a 192-bit interface and ranks among the best graphics cards; the cut-down prototype likely would have changed our feelings about the design.

Before the RTX 40-series debut, leaks suggested that Nvidia was looking to build its next-generation RTX 4070 with just 10GB of VRAM. Such leaks were perhaps true, or at least Nvidia was considering it, as we now have evidence that Nvidia actually built pre-production samples of the RTX 4070 in 10GB trim.

A GPU-Z screenshot of the prototype 4070's specs reveals how Nvidia created this original 4070 variant. Nvidia cut the memory width to just 160 bits wide, with five 2GB memory ICs providing 10GB capacity. This was done by simply disabling one of the six 32-bit memory controllers on the AD104 die. As a result, memory bandwidth and capacity were restricted, giving the prototype 4070 10GB card just 420 GB/s of memory bandwidth — 16.7% less than the eventual RTX 4070 that came out in April 2023.

Interestingly, Nvidia apparently hoped to compensate for the reduced VRAM and bandwidth by providing the GPU more CUDA cores. The RTX 4070 prototype has 7,168 CUDA cores — 1,280 more cores than the vanilla RTX 4070, and the same number as the RTX 4070 Super. Note also that the prototype has a PCB with eight memory solder locations, likely using a tweaked version of an AD103 (RTX 4080) board.

We have to say we're very happy that Nvidia did not implement this suboptimal memory configuration on the real RTX 4070. Even though this prototype model had noticeably more CUDA cores than the outgoing trim, the memory subsystem would have been problematic. There were also rumors that it would have slower GDDR6 memory rather than GDDR6X, though there's no indication of L2 cache size — a larger cache could have potentially mitigated some of the bandwidth impact.

The RTX 3080 10GB has been the perfect candidate for testing 10GB VRAM capacities in modern titles. We've found that 10GB of VRAM, on Nvidia GPUs in particular, is already showing noteworthy limitations in the latest titles. Horizon Forbidden West is a prime example of where 10GB is starting to show the same limitations as 8GB GPUs. In that game, at 1080p and 1440p, the RTX 3080 10GB does fine, but it completely falters at 4k, where 11GB of VRAM is required to run the game smoothly on the highest settings.

12GB of VRAM is now the lowest VRAM capacity we can realistically say modern titles (besides one or two exceptions) require to run at max settings, and we are glad Nvidia changed its mind to reflect this reality in the RTX 4070. In our testing, the RTX 4070 and its refreshed counterpart, the RTX 4070 Super, don't have any serious problems in modern games and work well at 1080p, 1440p, and 4K. That's chiefly thanks to both GPUs coming with 12GB of video memory.

In a perfect world, we wish the RTX 4070 series had 16GB of VRAM, as that would give them both some extra wiggle room (i.e. "future-proofing"). We've said in the past than an extra 4GB at most levels would have made the RTX 40-series Ada Lovelace GPUs more palatable. But we can appreciate that Nvidia didn't go the other way around with an RTX 4070 10GB. If that happened, 4070 users would be in a similar situation to RTX 3060 Ti, 3070, and 3070 Ti owners who have been struggling with memory capacity issues for years thanks to Nvidia low-balling those GPUs with just 8GB of memory.