There is a rumor about 48GB version of RTX3090 replacing the original Titan RTX with a similar price. That would make sense for Deep Learning workloads as 24GB is already too small for attention-based models.

It's 50% in ray-tracing performance, not raw performance, and it doesn't have big memory bandwidth improvements. It's not the Titan successor, it's the 2080Ti.

You can see this because the chip name is GA102, the 2 at the end indicates that this is a cut-down chip.

All previous generation Titan models have been cut down chips as well TU102, GV102, etc.

With one exception: the Titan V card that I mentioned which comes with the GV100 chip. But as I mentioned, that one targets a very different market segment with HBM memory, which the "RTX 3090" obviously doesn't target, since otherwise it would come with HBM2 memory like all the GA100 products.

>All previous generation Titan models have been cut down chips as well TU102, GV102, etc.

>With one exception:

This is simply wrong. The majority of Titans have been full chips: Titan, Titan Z and Titan Black as well as the Titan V and Titan X.

The other Titans were about as fast as manufacturer-overclocked 80Ti models.

Indeed.

The naming is super confusing. The 100 versions are essentially HPC chips on a PCI-express board, while 10x are completely different products.