NVIDIA GeForce RTX 3070 Ti based on the GA102 in the works?
According to the leaker which accurately predicted the GeForce RTX 30 series to date, there will be another graphics card based on the GA102 GPU. Previously there were rumors of the GeForce RTX 3070 16GB, which would be based on the full GA104 GPU, but that was allegedly scrapped. NVIDIA might instead offer a cut-down GA102 with 7424 CUDA cores as what could be the GeForce RTX 3070 Ti, or possibly the GeForce RTX 3070 SUPER, depending on how NVIDIA wants to play their cards this generation.
The 7424 CUDA core count points to a 58 SM configuration, 10 less than what the GeForce RTX 3080 offers. That would pack 232 Tensor cores and 58 RT cores, which is a nice bump up from the GeForce RTX 3070. In fact, this new GeForce RTX 3070 Ti (or SUPER) will sit quite well in between the GeForce RTX 3070 and GeForce RTX 3080, with 10 less SMs than the GeForce RTX 3080, and 12 more than the GeForce RTX 3070. On top of a higher core count, the GeForce RTX 3070 SUPER (or Ti) will also be offering a wider 320-bit memory bus, which might be used to feed 10GB of GDDR6X.
There are a number of reasons why NVIDIA might want to release a card in between the GeForce RTX 3070 and GeForce RTX 3080. Between these two cards, there’s a $100 (~RM416) gap, which is probably where the GeForce RTX 3070 Ti (or SUPER) will fit into. It hints that at least one of AMD’s RDNA 2 cards will fit into this price bracket, while offering quite a bit more performance than the GeForce RTX 3070, thus the need for NVIDIA to respond with this cut-down GPU at around $599 (~RM2490).
Another possible reason might be terrible yields for the GA102 GPU. NVIDIA has to bin the GA102 dies manufactured by Samsung to determine whether they can be sold as the GeForce RTX 3090 or GeForce RTX 3080, and if yields are very poor, there might a buttload of defective chips that cannot even meet the specifications of the GeForce RTX 3080. This might have forced NVIDIA to create a new SKU just to clear off the defective dies as a lower bin.
The full GA102 sports a total of 10752 CUDA cores, which does give NVIDIA quite a bit of room to play around with. Even the top-of-the-line GeForce RTX 3090 is essentially a cut-down GA102, as it only offers 82 out of the full 84 SMs that the GA102 GPU touts. For now, the only way you can get the full GA102 GPU is by getting the NVIDIA A40 and A6000 GPUs, which are targeted towards server and workstation segment, and will inevitably come with an even bigger premium than what the GeForce RTX 3090 has over the GeForce RTX 3080.
Pokdepinion: A GA102 for $599 must not sit well with the bean counters… Probably that’s why NVIDIA will work with TSMC next year.