With NVIDIA still not providing any open-source support or technical documentation for their graphics processors, for those in the open-source community who seek to use their GeForce 400/500 "Fermi" GPUs without NVIDIA's binary driver, they are left to use the reverse-engineered, community-created Nouveau driver. Fortunately, the support for the NVIDIA Fermi GPUs is coming along at a respectable pace -- with even working OpenGL acceleration -- considering that NVIDIA is providing no support at all. In this article are the first benchmarks of this experimental GeForce 400/500 "Nouveau NVC0" driver versus NVIDIA's official proprietary driver.
The GeForce 400 series launched nearly one year ago (April 2010) and just days after the GeForce GTX 470 and GeForce GTX 480 appearing, independent work began on bringing up support for this hardware in the Nouveau driver as well as within the PSCNV Nouveau driver fork. PathScale, the compiler company behind the PSCNV fork that's particularly interested in bringing up OpenCL/Compute support, even offered free Fermi graphics cards to developers.
Just months after the first NVIDIA GPUs appearing that were based upon the Fermi architecture, pushed into the Linux 2.6.36 kernel was NVIDIA Fermi KMS support. This was support in the Nouveau DRM (Direct Rendering Manager) driver for kernel mode-setting (KMS) for the GeForce 400 graphics cards out there at the time. However, this was just mode-setting support and there was no kernel or user-space support yet for 2D/3D/video acceleration. Not long after that, however, 2D and 3D began working. In October, PSCNV proclaimed stable 2D support for Fermi while 3D was a work-in-progress.
In January for the Linux 2.6.38 kernel was when Nouveau Fermi acceleration was merged for 2D/3D/X-Video. With the Linux 2.6.38 kernel, however, you need to use non-free firmware/microcode to initialize the GPU. This is still an unfortunate requirement as of today. If you load the latest Linux kernel on a NVIDIA GeForce 400/500 "Fermi" GPU, there will be kernel mode-setting support, but without any acceleration. There is no "free" Nouveau Fermi microcode/firmware that is redistributed with the Linux kernel and the Nouveau DRM cannot provide acceleration without the firmware.
This used to be a problem originally for earlier generations (GeForce 6 series and newer) of NVIDIA GPUs on Nouveau, for those that remember the ctx_voodoo controversy, but this had changed in late 2009 when a free firmware was created and then in early 2010 when it became possible for Nouveau to generate its own firmware (the DRM driver building its own "ctxprogs" and state internally). This should eventually be the case for Fermi too, but it will not be until the Linux 2.6.40 kernel or later where this changes. For now you may have kernel mode-setting "out of the box", but not acceleration.
The firmware needed by the Nouveau Fermi/NVC0 driver is not hosted publicly either, but interested users need to extract the firmware themselves. Using MMIO-Trace for tracing memory mapped I/O access within the Linux kernel when first using the NVIDIA binary driver does this and from there the binary files can be assembled. Basically you load the memory mapped I/O tracing support (prior to loading Nouveau, the NVIDIA blob, or anything else poking the GPU), load a recent release of NVIDIA's binary display driver that supports the given GPU, start X.Org Server so the GPU is initialized, and then from the recorded MMIO-Trace dump the firmware can be extracted. There are four "fuc" binary files needed for Nouveau support at this time (fuc41ad, fuc409d, fuc41ac, and fuc409c). It is a relatively easy process for an experienced Linux user and can be done in a few minutes, after which you are free from having to use the NVIDIA binary display driver. As of a few days ago, a brief how-to guide was posted to the Nouveau Wiki to ease this process.