At the GPU Technology Conference, Nvidia revealed its latest roadmap for the upcoming GPU architectures. We show you the way where the company plans to go in the upcoming couple of years.
Nvidia Public GPU Roadmap now shows key features on Maxwell and Volta
The roadmap in question clearly shows Maxwell arriving in 2014, one year behind the original schedule. One of key features of this architecture will be Unified Virtual Memory, expanding from the shackles of 1, 2 (and soon, 4 Gbit) GDDR5 memory chips to addressing the memory. However, what might surprise you is that Nvidia is not referring to the standard addressing of the Host CPU memory but through Fusion-io Solid State Storage as well. The company signed a technology partnership agreement with Fusion-io and they’re both looking into optimizing the future GPU architectures. The way how future GPUs (whole line up – Consumer, Professional and Compute) are going to operate is the minimal amount of on-board memory (SRAM Cache), significant amount of GDDR/SDRAM memory as close to GPU as possible and then tertiary memory subsystem in the form of either CPU host memory or Fusion-IO cache/work memory.
We also learned that Kepler architecture is not over with the new and products, with one potential product being a Tesla "K30", a GK110 based chip with 4 Gbit chips, enabling a total capacity of 12GB GDDR5. This news is certainly welcomed by the scientists around the world, which cannot wait for Maxwell and Volta.
Meet Nvidia Volta – Waving good bye to traditional memory topology, Volta will feature yet-unknown amount of Stacked DRAM
Coming in 2015-2017 timeframe (realistically, 2016), the Volta GPU brings the biggest change in the GPU architecture since its introduction in 1999 with the original GeForce GPU (NV10). This is the first confirmed part in the semiconductor industry to use Stacked DRAM technology. We wrote extensively about the technology (Hybrid Memory Cube: The Future of Exascale Computing, Micron & Samsung form Stacked DRAM Consortium) in the past and now it turns out that Nvidia Volta could be the first out the door with the technology.
Apparently, Nvidia was not happy with the performance and density characteristics of GDDR6 (set to debut in 2014 with AMD Volcanic Islands and Nvidia Maxwell), and the company decided to employ Stacked DRAM in order to achieve incredible bandwidth of 1TB/s. Every Stacked DRAM module currently achieves around 128GB/s, which means Volta needs to have at least six memory sets in order to achieve desired figure. Jen-Hsun Huang, CEO and Co-Founder of Nvidia also stated that the way how GPUs will be packed in the future will be significantly different than today, with Volta GPU being surrounded by what it seems to be six stacked DRAMs.
Before you go on overboard sharing of the picture above, bear in mind that this is just a mock-up, with i500 die shots "acting" as Stacked DRAM, and the GPU in question is GK110 (Volta won’t tape out before late next year).
All in all, the GPU architectures of tomorrow look very appealing. Bear in mind that Volta is probably the GPU behind the two of world’s first Exaflops super-computers.