The hot rumor to start the week kicks off with NVIDIA's upcoming next-gen Volta GPU architecture, which started off on the Beyond3D forums that NVIDIA will be using a custom 12nm process from TSMC. Remember this is all speculation and rumors, but we should have better clarification at GTC 2017 in May from NVIDIA direct.
NVIDIA shifting to 12nm is an interesting move, as the 10nm node is for SoC node - which is why Qualcomm is making its new Snapdragon 835 processor on the 10nm node, while 7nm isn't far away. The shift from to 12nm provides NVIDIA with more performance over the 16nm process, which is always welcomed.
GTX 20 series = Pascal Refresh
We wrote about the GeForce GTX 2080 and the entire GTX 20 stack of products, which will roll out as a refresh of the Pascal architecture. We shouldn't expect too much new from the GTX 20 series, apart from 10-15% additional performance, potentially lower temps and more overclocking headroom.
There won't be a shift to GDDR6 or HBM2 on the Pascal refresh, I think NVIDIA will leave the new RAM technologies for the Volta architecture. If the Pascal-based side of things did receive HBM2 on the consumer side (remember, the Pascal-based Tesla P100 is powered by HBM2) - it won't be until 2018 or later.
GDDR6 & HBM2 on Volta
This is where things get exciting. NVIDIA will most likely reserve HBM2 for their higher-end Volta-powered GeForce graphics cards, while GDDR6 will fill the gaps in between the highest-end GTX 2080 and GTX 2080 Ti - and the new series of cards. This in itself is a story - will Volta be part of the GTX 20 series, a new Titan series, or something completely new?
GDDR6 provides 16Gbps of bandwidth, up from the 10Gbps on GDDR5X (which is used on the GTX 1080 and Titan X), and up from the 7-8Gbps on GDDR5. It's a massive improvement, providing 60% more memory bandwidth over GDDR5X, and much more compared to GDDR5.
GeForce GTX 30 Series - What To Expect
What can we expect from NVIDIA's next-gen GTX 30 series cards? Well, we should see three different Volta GPUs released in 2018 - the GV104, GV102, and GV110. NVIDIA will use the GV104 in its slightly higher than mainstream segment at $400 or so, where we should see the GeForce GTX 3070 and GTX 3080 powered by the GV104 core, with up to 16GB of GDDR6 RAM on a 256-bit memory bus.
As for the higher-end GV102-based product, we could expect up to 32GB of GDDR6 or HBM2 on these cards at the full 16Gbps bandwidth providing 512GB/sec of memory bandwidth. This could scale much higher, especially if HBM2 is thrown into the mix, we could see up to 1TB/sec (1024GB/sec) of memory bandwidth.
The GV102 core will be the Titan X successor, which is why I have predicted we will see up to 32GB of RAM being offered. Sure, it sounds crazy now - but Volta is a 2018 product, and GPUs are going to be getting much more complicated, and much more technologically advanced in the next 12-24 months.