In this image, Tesla is referring to G80, Fermi is obviously the GPUs we have today in the GTX and GTS 400 series of graphics cards but Kepler and Maxwell are new to the scene.
Kepler is due in 2011 and will be built on a 28nm process technology, though whether that is TSMC or GlobalFoundries is yet to be seen. Huang said that by the time of release the company will have invested about $2 billion in the design. Little information was given on performance, but Huang was adamant they there were targeting performance-per-watt, and that is what the scale on the left hand side represents. Kelper will have apparently have 4x the performance per watt in terms of double precision GFLOPS.
That is an important note as we don’t expect the raw performance of the GPU in real-world applications and games to see that much of an increase in just a year. Double precision computing performance is more about high-performance and professional computing but is basically not used in any gaming environments.
But, if we base a Kepler GPU on a ~200 watt power envelope, we hope that this will still equate to at least a factor of 2x gaming performance compared to the results we see from Fermi today.
Maxwell is the following GPU architecture that is schedule for a 2013 release and will be built on a 22nm process technology. Compared to Fermi, Huang stated that Maxwell will bring a 10x performance per watt increase, again in double precision GFLOPS.
NVIDIA stated that with these new architectures they bring new features as well including things like pre-emption, virtual memory support, etc.
When asked about the move from annual architectures to bi-annual releases (as the Kepler/Maxwell roadmap indicates) Huang stated that they see little value in releasing new GPU architectures without also having a node process technology advancement. This means that as the progression in process technology advances, so will NVIDIA’s GPU plans. Kepler is targeting 28nm and Maxwell 22nm. Huang also commented that is node process technology advances faster they may alter their plans in the future. Mid-architecture “refreshes” are still planned in between these releases so the GeForce group could still plan on annual products, etc.
This does present a little more risk for NVIDIA though – combining a new architecture with a new process has been known to cause some significant issues every once in a while. Ask ATI with the Evergreen 40nm GPU or AMD with the Llano 32nm APU.
We still have concerns about NVIDIA quoting “performance per watt” rather than just raw performance; it leaves open the idea that performance will remain mostly stable while power efficiency will come down. Huang stated several times that “performance per watt IS peformance” because transistors are free moving forward, but power is not. True enough, but gamers will want to get as much performance as they can in the 200 watt envelope of the PEG slot and aren’t as interested in getting a GTX 480 at 50 watts. That feat would be impressive and is necessary for many other markets and areas, but let’s hope the performance progression rate increases as well.