At the company’s on-going GPU Technology Conference in San Jose, NVIDIA unveiled a couple of roadmaps that are well-worth paying attention to. On the GPU side, the company discussed “Volta”, the successor to “Maxwell”, which is expected to launch at some point in 2014.
During his keynote, NVIDIA CEO Jen-Hsun Huang listed the major features that have come with recent architecture launches. Tesla? CUDA, of course. Fermi? FP64 operations. Kepler? Dynamic Parallelism. The upcoming Maxwell will feature Unified Virtual Memory, which aims to allow both the CPU and GPU to use each other’s memory – potentially a major benefit for increased efficiency. Finally, there’s Volta, the first GPU architecture that will utilize stacked DRAM.
Stacked DRAM is an interesting beast. Because signals crossing the plane of the board is naturally inefficient, a stacked DRAM design would reduce the distance that the signal would have to travel. A real-world comparison would be having to walk across a schoolyard to get to another building versus simply walking up some stairs.
Because of this design, NVIDIA expects the memory bandwidth to triple (1TB/s!) – mind-blowing to say the least. Especially for the HPC market, that could prove to be quite the boon. For gaming, we’ll have to wait and see.
Aaand, that was about all the Volta speak that Jen-Hsun had to offer. Also discussed was the next Tegra chip (5, likely), which brings CUDA into the mix. This is obviously a big thing for NVIDIA, and could also be for consumers. This would be made possible due to the implementation of a Kepler processor, one that fully supports CUDA 5 and OpenGL 4.3. Computation aside, just picture the gaming potential!
I won’t re-write everything that happened at the event, but I highly recommend checking out NVIDIA’s own liveblog of the keynote.