Last week, we reported about a finding that Nvidia was using its cut-down GA104 graphics processing units (GPUs) for GeForce RTX 3060 graphics cards that are supposed to be based on the GA106 GPU. The 'new' boards have the same performance characteristics and power consumption, as the 'old' ones. This week two suppliers of graphics cards — Gainward and Galax — listed appropriate products, which is an indicator that such GA104-powered GeForce RTX 3060 boards are coming to retail.
Only in China. For Now
For now, Gainward and Galax only list GA104-150-based GeForce RTX 3060 graphics cards in China, so this version of the product may end up as a yet another exclusive offering for the country. Meanwhile, previously Nvidia switched GPUs on the GeForce RTX 1650 and on at least one GeForce RTX 2060 model, so it would not be a surprise if Nvidia decides to use the GA104-150 for GeForce RTX 3060 boards sold in other parts of the world.
Using a cut-down GA104 GPU instead of the GA106 for the GeForce RTX 3060 only makes sense if the silicon is not good enough to power GeForce RTX 3060 Ti/3070/3070 Ti graphics cards since it has too many defects. The GA104 is a pretty large 392 mm² chip that packs 17.4 billion transistors and up to 6144 CUDA cores that is rather expensive to make. By contrast, the GA106 is considerably cheaper, it is 276 mm² and integrates 13.25 billion transistors as well as up to 3584 CUDA cores.
Normally, chip developers like AMD and Nvidia try to sell batches of chips with defects (but which can still work perfectly with a few non-functional units deactivated) through OEMs and/or on specific markets via select vendors since it is not easy to support a huge number of SKUs across the world.
Yet, if a company has a significant number of dies that cannot work in proper configurations because of defects, they have to either invent new models or repurpose silicon for lower-end SKUs that are available globally. It remains to be seen whether Nvidia has enough GA104 chips that cannot be used even for GeForce RTX 3060 Ti boards (that carry GPUs with 4864 CUDA cores) and the company will have to sell them outside of China.
Gainward has two (or four, depending how you count) GeForce RTX 3060 12GB graphics cards that can be based on either GA106-300, GA106-302, or GA104-150: the GeForce RTX OC/OCG with a massive triple-slot cooling system with three fans and the GeForce RTX 3060 DU/DUG with a dual-fan dual-slot cooler, as discovered by @momomo_us. Although Gainward does not indicate this explicitly, it looks like the letter 'G' in the part number points to usage of the GA104-150 GPU.
Galax, which is Gainward's close relative, has four (or eight) GeForce RTX 3060 12GB add-in boards (AIBs) that may be based either on the GA104 or the GA106: the GeForce RTX 3060 Gamer MAX OC[FG], the GeForce RTX 3060 Golden Master MAX OC[FG], the GeForce RTX 3060 Xing Yao MAX OC Forever Edition [FG], and the GeForce RTX 3060 MAX Xing Yao [FG], according to @momomo_us. In Galax's case, it looks like the letters FG in the part number indicate that the board is powered by the GA104, yet we cannot be 100% sure since the vendor does not specify this clearly.
The graphics boards from Gainward and Galax use large printed circuit boards (PCBs), yet we do not know whether they use the same PCBs both for GA106 and GA104-based GeForce RTX 3060 12GB graphics cards. Meanwhile, if PCBs are the same, this would mean that Nvidia's GA104 and GA106 are pin-to-pin compatible. The good news is that at least it looks like GA106 and GA104-powered GeForce RTX 3060 AIBs have different part numbers.
A Blessing in Disguise?
While for now a GeForce RTX 3060 with a GA104-150 GPU is hard to find, there might a motive for customers to get these boards.
Since the GA104 is considerably bigger than the GA106, yet has many unused parts, it is easier to cool it down to normal temperatures, which means that these parts could be quite good overclockers when compared to regular GeForce RTX 3060 cards carrying the GA106.
Still, an overclocker has to keep in mind that overclockability also depends on PCB design and power delivery. Meanwhile, GeForce RTX 3060 AIBs are designed to be inexpensive to build, so setting GPU frequency records was not a goal that engineers intended to achieve with these products.