Kickstarter For Open Source GPU

I don't understand why their first goal is so pricey for what it delivers, but then it seems to go the other way and twice the cost rewards with a lot more than twice the work (just guessing though :-( Maybe they have some agreements in place with some of the developers to pay for releasing some of the code as open source?

That's a very good question.  Since they already have the 2D functionality working and they claim that there is no risk for the first target, what's the $200K for?

I believe it would be for Manufacturing the Chip.

What they've done up to this point is mainly Software.

R_Phoenix wrote:

 

I believe it would be for Manufacturing the Chip.

 

What they've done up to this point is mainly Software.

I see your wink, of course.

Just in case others miss it, $200k wouldn't come even close for a project involving delivering a physical device.  They explain clearly that the project entails delivering only Verilog code and related development materials:

Silicon Spectrum writes:

 

Our deliverables are Verilog source code and testbenches.

So it's an interesting question where the money will be spent for the (already nearly finished?) initial goal.  The other interesting question is about their expectations regarding the number of backers, since the number of people for whom Verilog is a worthwhile personal acquisition is not really very large, at least not by the standards of most Kickstarters.

I hate to say it but this looks like an attempt to get $200k for some Verilog they can't sell any other way.

They are incredibly vague (unless I missed something) about the size of FPGA and the performance that will be achieved.

It's also not clear if there is any comitment to open sourcing the drivers for Linux or Windows.

Any serious 3D graphics accelerator needs GB of RAM which will be off chip - it takes a lot more than the basic IP to turn that kind of technology into  a working board so I'm not at all sure who the later stages of the project are aimed at.

MK

Unfortunately I have to agree with you.  Although I would like to advocate strongly on behalf of OSHW enthusiasts and encourage them to do a lot more to bring hardware under their control instead of just taking what they're given and having to like it, I don't see how this project achieves that.

Equally importantly, on the Kickstarter front it doesn't deliver value for money, in that the number of backers for a Verilog deliverable is likely to be extremely low, and therefore the amount pledged per person would have to be extremely high to meet the target, which mades no sense on value for money grounds for an open HDL design.

The days of paying a lot of money for code are long gone, and that applies to HDL just as much as to code for CPUs.  Some people just haven't noticed the sea change yet, and this is a particularly bad case of it.

Michael Kellett wrote:

 

I hate to say it but this looks like an attempt to get $200k for some Verilog they can't sell any other way.

Seems that way. In addition to the general vagueness, two things that caught my eye were the PCI interface, not PCIe or the type of interfaces common for SoC's, and Direct3D 7/8.  So the design seems rather out of date already.

Then stuff like bump mapping being in the $800,000 goal and the early 90's style screenshot just adds to the amazement.

Performance wise, did you watch the video ?  If it's representative of expected results, well...

Morgaine Dinova wrote:

 

PS. I've been wondering whether the Parallella board wouldn't be a nice implementation platform for the programmable pipeline of a modern open-source GPU.  After all, the Epiphany device provides hardware parallelism with fast floating point for the core compute engines of 16 to 64 shaders without having to implement CPUs in Verilog.  The FPGA would then be left to do only the fastest parts of the overall task instead of implementing the programmable shaders that CPUs are best at.

I think that'll depend a lot on whether the Epiphany provides the type of operations typically used in GPU's for shaders or stuff like h.264 encode/decode and is fast enough at those particular things.

While I'm sure it would be possible, is there any point if it ends up being beaten in performance by a $50 mass produced ati/nvidia card ?

Getting decent open source drivers is a worthy goal, but I don't know how feasible the idea is. Most of the problems with GPU's today is to do with the technology/algorithms being wrapped up in patents and intellectual property law suits, not the ability to create the implementation or drivers.

Even the premise of designing an OSHW GPU targeting an FPGA is rather questionable, given how no FPGA in existence is programmable with an open source toolchain.  Although they're also proposing to license it for ASIC production, it's debatable whether a primitive design which only committed OSHW fans could love would ever be chosen by such a manufacturer given it's likely meagre sales.  The business plan doesn't seem very convincing, no matter how much I'd love to see an OSHW GPU.

Althought their bailliwick is Verilog, I wonder if an OSHW GPU wouldn't be better implemented in a completely different way, a way that open source software people would find more suited to their skills.  It seems a bit silly to implement the shaders of the programmable pipeline in an FPGA when the Epiphany device is already available and provides 16 cores of ASIC efficiency ready and waiting to run shader processing in a novel way.  Nobody said that GPUs can only have the design preferred by nVidia and AMD.

selsinork:  Our replies overlapped, but I'll just say that an OSHW GPU doesn't have to look anything like the commercial ones.  3D graphics continues to evolve, and raytracing scenes or voxel processing isn't necessarily handled optimally by the current generation of graphics cards.

Morgaine Dinova wrote:

 

Nobody said that GPUs can only have the design preferred by nVidia and AMD.

 

selsinork:  Our replies overlapped, but I'll just say that an OSHW GPU doesn't have to look anything like the commercial ones.  3D graphics continues to evolve, and raytracing scenes or voxel processing isn't necessarily handled optimally by the current generation of graphics cards.

Agreed.  Ultimately though there's a couple of things a GPU needs to be able to do. For the sake of argument lets say that boils down to H.264 and OpenGL. Actual implementation details aside, you're going to have to deal with accelerating those API's in hardware.  Do some random thing of your own that no software supports and you have a product with no market and are likely doomed to failure.

So whether you implement OpenGL & h264 in the same way as everyone else, or do it with an array of Ephiphany cores and some novel approach to the problem is almost irrelevant.  It all comes down to whether the competition can do the same task faster and/or cheaper through the same API. If they can do either then your chance of success will be reduced, if they can do both...

I'm all for new and interesting ways to do things, especially if it's done in an open way that benefits everyone by giving us access to the source.  Still, if it's not competitive you become a follower not a leader and nobody will care other than a few die hard OS evangelists.