It turns out that NVIDIA originally planned a version of RTX 4070 with a rather paltry amounts of VRAM, but (thankfully) that wasn’t to be. According to @harukaze5719 on X/Twitter, One such card has emerged on Chinese marketplace 闲鱼 (Goofish), which comes with vastly different specs than what is available today.
RTX 4070 10GB?
The GPU in question is an unusual one – the silicon variant here, labeled AD104-275, has not been used before, though leaks from almost 2 years ago does indicate that NVIDIA had ditched this variant before the card was officially launched, with less CUDA cores on tap in the retail model. The AD104-275 silicon has the same amount of CUDA cores (7,168) as the RTX 4070 SUPER today, which uses AD104-350 variant; meanwhile, the vanilla RTX 4070 uses AD104-250 silicon that only comes with 5,888 CUDA cores.
However, both RTX 4070 variants does come with 12GB of GDDR6X VRAM via 192-bit bus, whereas this unreleased variant comes with just 10GB on tap, with a narrower 160-bit bus. This means memory bandwidth will be worse, and this can easily impact framerates as games demand more video memory and its bandwidth as resolution goes up. However, based on the GPU-Z readings, it looks like this variant is supposed to have a significantly higher GPU clock, likely to compensate for the performance penalty caused by the lack of VRAM bandwidth.
The circuit board, which looks like intended to be a Founders Edition design, also appears to have one 12VHPWR connector. It’s worth noting that while the RTX 4070 comes with 12VHPWR connector at launch, AIB variants comes with traditional PCIe 8-pin connectors as well. That said, the TGP rating of this card remains unknown, so we don’t know what kind of performance this card would’ve targeted originally.
Pokdepinion: Good thing we have 12GB variants today. 10GB is just… nope.