Adapteva Parallella 18-core Motherboard

This is by far my favorite development board I've ever had. So the board contains a total of 18 cores, but 16 of them are located on a separate ASIC and it is called the Epiphany. It is essentially a co-processor with 16-cores and its own instruction set architecture. The other two cores reside on the Xilinx Zynq FPGA and they are ARM processors.The reason why I love this board so much is because of how much you can learn about building a system from the ground up including custom hardware, drivers, compilers, etc. Everything is open source except for the the Epiphany co-processor itself. The board runs a fork of Ubuntu called Parabuntu which contains drivers that allow the host CPU (ARM Cortex) to communicate with the Epiphany 16-core co-processor. The drivers move data between the host and the co-processor via their eLink protocol. The FPGA contains a synthesized component that handles this and the Linux driver interfaces with it. All of this is open source, so you can see how the driver is written, and even the Verilog hardware description language for the eLink components. Since Parallela doesn't consume all of the logic cells on the FPGA, you can even add your own custom hardware accelerators and synthesize them along with the Parallela components. In fact they have a library called "OH" for open hardware which simplifies creating additional hardware accelerators and there is tutorial on how to do it.So I love this board because you can see how parts of the hardware are built, the drivers are written, and the GCC compiler was forked to support the Epiphany's custom instruction set. That alone makes this a top notch board; however, I haven't even touched on the secret sauce.The Epiphany packs 16-cores with its own general purpose computing instruction set. So these cores are not as primitive as a CUDA core on an NVIDIA GPU. They are general purpose just like your Intel / AMD / ARM processor. You can either use C/C++ to compile code for the Epiphany or you can put your hacker hat on write assembly. Or even better, do both and inline the assembly where needed. It also supports Python called ePython by the person who worked on that project.These cores are general purpose, but the memory model is different compared to your typical dual or quad core Intel chip. The cores are connected via a mesh network which is optimized for write transactions. Each core has its own small chunk of local memory, can access the RAM if needed, and read / write between cores. This memory model is what allows this architecture to scale to 1024 cores with ease, but it also presents a fun challenge for developers. Knowing how many clock cycles a write or read transaction takes between cores makes think differently about how your algorithm should be implemented. Also, each core can run a completely different program which is pretty cool. The host processor simply copies the compiled ELF to the cores you want to run it on and then kicks them off running.There are also a few libraries available in addition to the eSDK that allow you tackle multi-core programming differently like OpenSHEM (shared memory) and COPRThreads (co-processor threads).This board is just the best all around, and I hope the community keeps it alive. I know Adapteva didn't hit critical market mass with this board, so Andreas the founder had to take a job at DARPA, but so far the community seems to keep going. Unfortunately we won't see the 1024 core Epiphany for ourselves.

I just don't get this thing. I tried to use it with OpenMP on Fortran programs mostly for the numerous DO loops in the code. I every case I tried this processor/OpenMP resulted in considerably slower speed than my dual core laptop regardless of the number of parallel channels used. In every tutorial (worth anything) I found on the web they compared the improvement in processing time of single channel use of the parallela to multichannel use on the parallella. A very slight improvement in speed was achieved this way but so what? Any dual core or better can outrun this board. They also claim you can learn parallel processing using this board. What does that mean? Buying a book on OpenMP or other parallel processing technique and downloading the software? I'm sorry I paid $100.00 for this thing.

This is a device that makes it possible to play with multi-computing on a budget. I used to work on large "Massively Parallel Processing" projects including OSF1/AD. The equipment we used were unique and cost mega-bucks at the time. No normal human could afford them and it was only because of my job that I got access to them. This allows anyone to work on a platform starting with 16 processors. I have not done a whole lot beyond booting and setting it up yet but it looks good so far. Therefore I will give it a high rating even though it could use a few improvements especially improved documentation.Nevertheless it is a great start and I wish Adapteva the best. My suggestion is that if you are buying one of these then also get:1) Be a developer who is comfortable with hardware and software and have a regular Linux box around. Ubuntu is nice but any Linux would do.2) Have at least one micro-sd card of 16 or better 32 gig size along with reader/writer. It must be *micro* not regular SD!3) Have a powered USB hub4) Have a micro-usb to USB adaptor5) Have a micro-hdmi to regular hdmi adaptor and a monitor able to display it.6) Have a serial RS232 to USB cable for debugging. Hopefully you won't need it but it is cheap and might as well have it.You might want to get a fan also but the newer ones with large heatsink will do OK without a fan. I'd still get the fan to be safe.If you get in trouble they got a great forum that will help you fix the problems but you have to be willing to work at it. This is not some out of the box product. You need to be willing to roll up your sleeves and dig in there. I spent 3 days before it booted properly for me. The people on their forum are very nice (I am there :-) and will help you get it working so have no fear.

I purchased four of these, along with a cooled enclosure, to act as my new desktop and research machine. It takes a bit of work and some good UNIX skills to get it set up and working correctly, but it is well worth the effort. These devices are excellent distributed/parallel computing research platforms. You can write parallelized algorithms in C or Python (easily). I've no experience with their support for other languages. Note that the forums on parallella dot org are extremely helpful, though quite technical.One other important thing to note is that even though they say you can run the board with just the heat sink, my monitoring of the chip temperatures showed the board quicky overheats, even when mounted vertically, as recommended. This computer requires, IMO, a case with a fan. Certainly, a fan is required if you want to enclose the boards or run more than one in a cluster.I am extremely happy with these computers, and I do recommend them. They really aren't for beginners, though. I would suggest you only buy these if you have excellent UNIX and system administration skills.