NVIDIA Reveals GPU Roadmap at GTC – Volta and Stacked DRAM ComingWed, Mar 20, 2013 - 1:12 PM
NVIDIA announced an updated GPU architecture roadmap at the 2013 GPU Technology Conference (GTC) that might interest you. NVIDIA’s current desktop graphics cards lineup uses the “Kepler” GPU architecture, but that will change next year when NVIDIA’s next-generation “Maxwell” GPUs are expected to debut. If NVIDIA’s slide is scaled correctly it looks like we can see double the performance-per-Watt with “Maxwell” versus what we are using today.
The slide also shows that Maxwell GPUs will feature unified virtual memory, which lets processors use the graphics card memory as system memory, for faster general-purpose performance of the entire system. After “Maxwell” comes “Volta” that uses stacked DRAM technology. The “Volta” GPU is named after Alessandro Volta, the Italian physicist who invented the battery two centuries ago. “I love that name Volta because it would suggest that it will be even more energy efficient,” explained NVIDIA CEO Jen-Hsun Huang. Notice there is no date on the X-axis, but we would expect this architecture to come out sometime in 2016.
Huang said that Volta is going to solve one of the biggest issues with GPUs today, which is access to memory bandwidth. Numerous companies are working on 3D stacked memory, so it is the future. Nvidia said they will be putting the stacked DRAM and the GPUs onto the same silicon substrate and inside of the same packaging before it welds that package to PCB. This sounds tricky, but if NVIDIA can get this to work it would be extremely fast. The image above shows an example with six stacked DRAM units plus a GPU in a single chip solution. Huang said that the “Volta” GPUs should have an aggregate of 1TB/sec of bandwidth into and out of the stacked DRAM. That is some insane speed. To put it into perspective, you can move an Blu-Ray movie through the memory in 1/50th of a second! You can watch the GPU roadmap announcement below for yourself.
“Next two GPUs from us: Maxwell is coming with unified virtual memory, which makes it possible for GPU operations to see the CPU memory and vice versa, so program-ability is easier. After that is Volta, which is even more energy efficient, which has a new technology called stacked DRAM.” Jen-Hsun Huang