Another Look At The Open-Source Nouveau With Fermi

Earlier this week was benchmarks of the NVIDIA GeForce GTX 460 "Fermi" with the open-source Nouveau driver. The reverse-engineered Nouveau support for the GeForce 400/500 series is incomplete and requires users to generate their own custom firmware before there is even 2D/3D/video acceleration support. The initial tests on the GeForce GTX 460 also yielded a disturbingly large performance difference between the open-source and closed-source NVIDIA drivers, where as with previous generations of NVIDIA GPUs the performance difference is more manageable. Here is another look at Nouveau for Fermi, but this time from a GeForce GTX 485M.

The tests earlier this week showed the binary NVIDIA driver being multiple times faster than the open-source Nouveau Gallium3D driver. The GeForce GTX 460 on Nouveau even ended up being slower than graphics cards of previous GeForce generations. It turns out that this support is due in large part due to the Nouveau driver not yet supporting re-clocking under Fermi. The Nouveau driver is stuck operating the graphics card at much lower core and memory clocks than the NVIDIA driver, which dynamically adjusts the frequencies based upon load via PowerMizer. Nouveau has begun supporting dynamic re-clocking and power management for previous generations of NVIDIA GPUs, but nothing is in place yet for Fermi.

Because of this incomplete support, the Nouveau driver is severely handicapped with Fermi. With Nouveau using the Linux 2.6.39 kernel, the open-source driver is running at whatever the default frequency of the GPU and video memory were set running when the kernel module was loaded. To force the Nouveau Fermi driver right now to run at full-speed, you would either need to manipulate the video BIOS or first load the NVIDIA binary driver, raising the clock speeds, and then reloading Linux (kexec -f) and then unload the NVIDIA driver and load the Nouveau DRM without rebooting the system. Obviously, this is far from ideal.

The current Nouveau support also lacks temperature and fan control support for Fermi.

Fortunately, the Nouveau developers are hoping to ship support for these critical features in the Linux 2.6.40 kernel.

In terms of the rendering problems reported earlier this week, it turns out that it's a mysterious bug impacting the Nouveau driver that's been referred to as "page jumping" where the virtual to physical memory address mapping is temporarily confused.

These comments were made by Nouveau developers in this Phoronix Forums thread. It was also mentioned that Nouveau developers have successfully reverse-engineered NVIDIA's CUDA (Compute Unified Device Architecture) from the binary driver. The Nouveau developers now know how to compile a CUDA program and execute it on the hardware. However, the CUDA/GPGPU/OpenCL focus is not as high for Nouveau developers (that's more the interest of the PCSNV fork) so there is not code yet in the open-source Nouveau driver for actually supporting open-source CUDA. There's also infrastructure work needed within Mesa / Gallium3D such as by the "Clover" OpenCL state tracker too.

The GeForce GTX 485M testing was done from a System76 Serval Professional notebook with an Intel Sandy Bridge CPU. Testing was done with Nouveau at its defaults and not forcing the higher clocks, so take the results as you wish. The GeForce GTX 485M results for Nouveau Git and the NVIDIA binary driver can be found on OpenBenchmarking.org.

There is also a CPU usage comparison and other metrics for Nouveau and the NVIDIA driver. This is done using the system_monitor module for the Phoronix Test Suite when setting the environmental variable MONITOR=all.

Long story short, the Nouveau GeForce 400/500 "Fermi" support still has a ways to mature. Still needing to generate your own firmware with there being no redistributable firmware (nor is the kernel driver able to generate itself) and hardware acceleration not being available without it, that's the most crippling aspect right now for most end-users. However, once you do have acceleration support, you are with a high-performance GPU that is stuck with reduced clock speeds as there is no dynamic re-clocking / power management support for now. The Linux 2.6.40 kernel will hopefully address this in the coming months, which should be an interesting release too for Intel users with more Sandy Bridge performance improvements.