Toward the end of the opening keynote of NVIDIA’s GPU Technology Conference, CEO Jen-Hsun Huang uncharacteristically revealed the company’s GPU roadmap for the next three years. In general, the GPU maker likes to keep such plans to itself until it’s fairly close to a product launch. In this case, though, Huang laid out the next two GPU iterations under development at NVIDIA. As far as I know, even the press and analysts had not been pre-briefed on this surprise news.
Keeping with the scientist naming theme, NVIDIA started with “Fermi” in 2009, the next CUDA GPUs are to be “Kepler” in 2011, followed by “Maxwell” in 2013. The Kepler GPUs will be on the 28nm process technology node, and the first products are scheduled show up in the second half of next year. That implies NVDIA is well into the development cycle, with Huang noting that the company will probably sink $2 billion or so into the R&D for this chip. Maxwell’s process node is not yet defined.
No other details were revealed about Kepler and Maxwell, with the exception of their targeted performance/watt numbers. Kepler will aim for 5 double precision gigaflops per watt and Maxwell will shoot for around 15 gigaflops per watt. That represents a 3-fold and 10-fold improvement compared to today’s Fermi GPUs. If we assume that the top-end Tesla parts will basically maintain the same 200-plus watt thermal envelope, that means the Kepler Teslas will deliver around 1.5 teraflops and the Maxwell Teslas will top out at around 15 teraflops (now you’re talking).
Apparently the fun doesn’t stop there. Huang suggested they would continue to evolve the feature set on the next-gen GPUs as well. Capabilities like thread preemption, virtual memory, and devising a way to relieve the CPU memory bottleneck are all on the table. And even though the GPU chip generations are on a two-year cadence, there will be mid-life upgrades available in the off years — so Fermi should be due for one any day now.
The unveilation of the roadmap might have been just a bit of PR theater for NVIDIA, but it more likely points to a shift in the way the company sees its new cloud server and HPC markets. In high performance computing, especially, customers need longer lead times so they can start planning for system deployments two to three years in advance. Making sure everyone sees the road ahead is probably going to become standard operating procedure for the GPU maker as it continues to pentrate markets once reserved only for CPUs.