While it's nice finally having access to FPGA's on the cloud, it's still a shame that there isn't much going on with Open source development on FPGA's, on either the toolset or libraries/modules. Though this is largely due to manufacturer hostility. Maybe wider access and adoption might finally do something about this situation.

Stanford has a research group specifically on this issue: https://aha.stanford.edu/

Yes, agree. Also, SDKs are quite expensive - while GPU vendors give theirs away for free. This might be one reason for relatively low adoption vs GPU.

I don't think that is it. The tools aren't open source free, but they are dollar free.

Xilinx is the leader in FPGAs and their tools are free. Lattice tools are also free. Not sure about Altera, who is #2.

I've worked multiple places that shipped products using only the FPGA vendor's free offerings. The FPGA vendors see the tools as a way to sell silicon.

Altera was acquired by Intel last year and they dropped the Altera brand - they are now Intel FPGA.

There is OpenCores the equivalent of Open Source for Verilog/VHDL http://opencores.org/.

I'm curious what workloads you would want to see run on the FPGAs that would justify a more open platform?

As it is, I think Amazon's approach for AMI Reserved Instance Marketplace [0] sales makes sense for the AFI packages -- I would think it would make sense for a vendor to develop, debug & maintain the accelerators I'm aware of.

[0] https://aws.amazon.com/ec2/instance-types/f1/

I've used FPGAs for image processing. While OpenCL does accelerate this, at the end of the day it's still largely General Purpose Computation on a GPU. I'd love to be able to install a suite of Special Purpose cores for image manipulation, for example. Especially one that works with imagemagick, so that all the various libs compiled against it would just work.

Is anyone doing open hardware FPGAs (I don't mean reverse engineering proprietary FPGAs, I mean implementing new silicon)? Recently a RISC-V company implemented a simple 32 bit CPU on the ancient TSMC 180nm process. NRE was apparently quite affordable and turnaround time was a few months. So perhaps a startup could work on implementing a 1990s-era PLD in the same process, just as a starting point to judge market demand.

If you're looking to learn an HDL on an FPGA, it's probably still cheaper and easier to just buy your own, maybe something like http://store.digilentinc.com/arty-board-artix-7-fpga-develop... ($99). And for that matter, you can also get started by just using a sim and no FPGA for free

If you don't care about the DSP units you can pick up a CPLD kit for ~$20 [1].

[1] http://www.digikey.com/product-detail/en/lattice-semiconduct...

You can also get a "real" -- albeit small -- FPGA from Lattice for $25:

http://www.digikey.com/product-detail/en/lattice-semiconduct...

The Arty is a much more capable board, though.

What do you like about the Arty?

I'm considering the parallella board, which is around the same price.

https://www.parallella.org/board/

Between the two I'd recommend the Arty without hesitation.

The Parallella is intended as a demonstration board for the Epiphany SOC. The $99 model has no GPIOs or JTAG interface, making it essentially unusable for FPGA development; the higher-end models will still be awkward to work with.

The Arty is intended for FPGA development, and is much better suited for it -- built-in JTAG controller, lots of GPIOs on easily accessible connectors, user I/Os, etc.

Ok - I guess I'll mention Papilio - basically a Xilinx FPGA on a board. The Spartan 3 version is as cheap as $38 and the Spartan 6 with an HDMI phy is $75.

http://papilio.cc

I agree, the devkit boards are very reasonable now days. I think the part that should be in the cloud is the SDK's. The synthesis is the most computationally expensive part and you could cut compile times down by an order of magnitude if you could run you dev on big iron.

Thanks for the link!

Here's the discussion I found from nearly two years go when looking for more about this topic:

https://news.ycombinator.com/item?id=9863475

Is there one that works with Mac? I have an old one in a box that I haven't succeeded in getting running and I don't anything other than a MBP to run on (and I really don't want to do boot camp...)

I have actually been looking into this the past couple days:

* The opensource toolchain of icestorm [1] is the only thing running natively on mac.

* Most people run the vendor SDK on virtualbox under windows.

* I heard also of people using wine/playonmac in rare cases.

* I have tried installing the Icecube2 SDK from Lattice virtualized under Linux, because I didn't want to sacrifice the ~30GB discspace for a virtualized Windows. The installation was painful to say the least and I could not get it working on a recent Ubuntu versions because their license manager (Yes, the software is still free) has trouble with the network interface names (expects an eth0 interface).

[1] http://www.clifford.at/icestorm/

For reference (30 November): https://aws.amazon.com/blogs/aws/developer-preview-ec2-insta...

FPGA newbie here; How does this relate to toolsets that Cadence, Synopsys and Mentor Graphics and others make for chip design? Would you use tools from these companies to design VHDL / Verilog for the 'FPGAs as a service' or is this at all related?

Yes, you use EDA tools to synthesize (compile) and debug VHDL/Verilog code for FPGAs, although Xilinx provides their own tools.

Those tools are very close to silicium level design tools. I guess the smartest application of FPGAs on EC2 would be for FPGA design tools, they are incredibly slow.

"Each FPGA contains approximately 2.5 million logic elements and approximately 6,800 Digital Signal Processing (DSP) engines."

Can an FPGA guru provide a real-world example of the kind of processing power this provides?

It's very difficult to relate to conventional CPUs because the model is so different; each of those DSP engines could be said to correspond to a lane in a SIMD instruction. However, they're usually set up for integer maths rather than floating point. So you could say that it's equivalent to 1,700 Pentiums running MMX instructions. But it's still limited to one or maybe two DRAM interfaces, so your data has to be very "local" and preferably streaming through the system to get that kind of throughput.

Just from experience. Bitcoin mining 10x from gpu to fpga so there's a lot of potential. But then again we aren't even using gpgpu in web apps.

I do :)

Perhaps we should rename them to CPGA (Cloud programmable) - Field programmable seems a bit archaic when they're cloud-based