Just one week after Intel’s IDF, it’s over to NVIDIA’s graphics card event, the GPU Technology Conference in San José. At this conference, centred on GPU computing and pro graphics, NVIDIA unveiled a roadmap, something that’s quite rare in the world of GPUs but which gives us the company’s technological horizon.
Following the Tesla (GT200, GeForce 200, ~Quadro x800, Tesla 10) and Fermi (GF100, GeForce 400, Quadro x000, Tesla 20) architectures, NVIDIA will introduce the Kepler architecture as of 2011. This will exploit the 28 nanometre fabrication process and NVIDIA is talking about a gain, in terms of performance per watt in double precision, of between 3 and 4x on Fermi. To reach this, we imagine that NVIDIA will double the number of processing units, increase clocks and above all make an effort in terms of energy consumption of the chip in order to avoid going off the scale.
Later, in 2013, Maxwell architecture will come online. Although NVIDIA is aiming towards the 22 nanometre fabrication process, this will of course depend on TSMC and GlobalFoundaries. Still in terms of performance per watt in double precision, NVIDIA is counting on trippling once again the yield achieved with Kepler. 1024 processing units for Kepler and then 2048 for Maxwell?
Moreover, NVIDIA CEO, Jen-Hsun Huang, also said that he was planning on introducing an intermediate development for each of these new architectures. There will therefore be an evolution of Fermi and of the GF100 before Kepler, in other words within the next 6 months. Note however that NVIDIA is presenting Fermi as a 2009 architecture and the schedule should therefore not be taken as gospel. It has to be said that in contrast to Intel and its tick tock model, NVIDIA does not control the fabrication process on which its products depend, which adds a significant risk factor into its programme.
Lastly, Jen-Hsun Huang said that the late arrival of Fermi and Tesla architecture cards was an anomaly that the company will be making efforts not to reproduce in the future. A whole range of GPUs derived from the Kepler architecture should therefore come onto the market in the 3 months following the arrival of the first.