Parallella Community

The GPU you referenced is $9,900 for the 3D core. This would be a single license and probably not source.

The purpose of the kickstarter is to open source our 2D/3D IP under LGPL.

If you want to play with a free 2D/3D core, you can try opencores. There is one there, but it is integer based and uses a shift and subtract approach for the 1/W calculation which means it will run extremely slow. The size is suspect as it is over 1/4 the size necessary to do everything needed for a real openGL pipeline.

The core we have on kickstarter is full floating point and field proven. It runs 1pixel/cycle for rendering, and it can setup about 4 million triangles/sec. (probably more, but I'd have to do the math.)

Parallella is designed more as an interface to the Epiphany than as a full-blown desktop computer. Although it can serve as a rather restricted barebones desktop machine, the resources are going to run out pretty fast if it's treated as something that it's not. For a start there is limited and non-expandable RAM and no SATA controller, so not having a GPU is the least of its worries. Some of these things can be remedied on a daughter card, but it'll require a lot of expansion to rival a standard x86 desktop in that role. Parallella is primarily a computation engine.

Because of the above, I'm not sure it's very sensible to use up the FPGA for a cut-down GPU, when it can be used as a computation accelerator to make the Epiphany array even more effective. If people want graphics, just run X11 apps on Parallella over Ethernet into an X11 server running on your desktop PC. After some optimization the Ethernet is expected to have a sustained throughout of around 700Mbps which is ample even for ambitious graphics. (And your PC's GPU is very likely to be many orders of magnitude faster than anything that could fit in the FPGA.)

I doubt that a restricted GPU makes much engineering sense in this architecture, especially when Parallella boards are clustered and the system becomes highly distributed --- the graphic relationships won't be symmetric, and having to plug a monitor into one specific card is a questionable idea. A network architecture is a much better fit I think.

Just my engineering assessment. If a restricted GPU solves some engineering requirements that I haven't considered, I'd like to hear them.

PS. I'm totally supportive of the idea that the open source community deserves its own GPU design --- that's a really awesome idea, and long needed. However, the Zynq's FPGA is entirely the wrong place for it architecturally, and the amount of 3D acceleration that would fit is unlikely to be satisfactory since even old graphics cards set quite high expectations.

PPS. There is one GPU configuration that I do think would be viable, and that is to use the entire Parallella as an accelerated GPU graphics card, employing all of the Epiphany's cores as the GPU's programmable shaders. That would stand a chance of offering reasonable 3D performance, and it would also be Ethernet-connected and hence inherently a capable X11 server. Designed that way, it could end up as the open source community's first open graphics card with effective 3D, and hence a major achievement.

We chose LGPL as we feel it should be treated like a compiler library. If you change the code for the GPU, the changes need to be released, but it can be used in closed source projects.

Some of the conversation is interesting. I would view our GPU as an extension for people to use in their projects. The idea would be to free up processor and other FPGA resources by providing something to accelerate local displays. By using the Epiphany chip to accelerate the graphics, I think you are loosing the power of using it to develop other applications.