From over at linux.com there's an update on the Parallela, a low cost open parallel computing platform.
However, I'm pleased to be able to report that backers of the Kickstarter campaign have now all received their rewards and the Parallella board has finally gone into general availability, with international distribution also having recently been put in place. And I think (hope!) that few would argue that what has been achieved over the past 18 months is nothing short of remarkable.
The platform itself is built around the Epiphany multicore chips developed by Adapteva, which offer a 16 or 64 core RISC processor array, paired with a Xilinx Zynq 7000 series SoC which has a Dual-core ARM A9 and supporting programmable logic and peripherals. It's in a small (credit card) form factor
It's available off the shelf through Amazon, or internationally through RS, in the 16 core version. RS currently quote at around the £80 to £160 mark depending on the specific platform ("server", "development" or "desktop" versions of the SBC). Amazon seem to be quoting $120-$250 for a similar set of models. I can't see current availability on the 64 core version, though.
There's an earlier linux.com background piece with more information, and the main product page has specifications and technical details, including a reference manual.
Brief Spec summary:
Zynq-Z7010 or Z7020 Dual-core ARM A9 CPU
16 or 64-core Epiphany Multicore Accelerator
1GB RAM
MicroSD Card
2x USB 2.0
4 general purpose expansion connectors
10/100/1000 Ethernet
HDMI port
Linux Operating System
54mm x 87mm form factor
(Score: 2) by cafebabe on Friday September 05 2014, @10:44PM
I was directed to the Epiphany architecture a while back and it is a fairly capable RISC/NUMA design. However, I didn't think I'd see them available at such low cost. This development means that people can play with a credit card supercomputer for about six times the cost of a Raspberry Pi. Although the 64 core version may be delayed, there remain many applications where it would be useful to connect dozens of shared memory instances via one Gigabit Ethernet switch.
1702845791×2
(Score: 2) by LoRdTAW on Friday September 05 2014, @11:21PM
Six times the cost of an RPi? The parallella costs $99.
(Score: 1) by khedoros on Friday September 05 2014, @11:31PM
(Score: 1) by barrahome on Friday September 05 2014, @10:51PM
This can be used as a normal computer?
(Score: 0) by Anonymous Coward on Friday September 05 2014, @11:07PM
Yeah I would like to know details too. I didn't see much skimming the links.
Like does each core have an MMU?
How is scheduling handled?
What is the memory bandwidth - both local to each core and through the interconnect?
(Score: 4, Insightful) by cafebabe on Friday September 05 2014, @11:50PM
The ARM cores have an MMU and use Linux for process scheduling. The Epiphany cores work more like Inmos Transputers or Connection Machine nodes. The Epiphany cores would be used with a library to provide some provide parallel processing abstractions, like Occam semaphores or flood filling. Code running on Epiphany cores would otherwise be bare metal to maximize performance.
Your application is almost guaranteed to maximize I/O channels. That's fairly standard with supercomputing. Regardless, I believe that the first generation of Epiphany cores allow data transfer at 1.6GB/s simultaneously each way (bi-directionally North, East, South and West). That may seem quite large. However, when these links are part of a 4*4 grid (or 8*8 or 32*32) and addresses are also sent over these channels, the data throughput for each core is significantly lower.
1702845791×2
(Score: 1) by barrahome on Saturday September 06 2014, @10:49PM
What about video and more ram!
(Score: 1) by khedoros on Friday September 05 2014, @11:15PM
(Score: 3, Informative) by cafebabe on Friday September 05 2014, @11:19PM
Yes. Imagine a dual-core Raspberry Pi with Gigabit Ethernet, an FPGA and a small supercomputer fabric. It is entirely possible to use it like Raspberry Pi and ignore the other stuff. If you want to use the more exotic features, the process is akin to GPU programming. Thankfully, the system is self-hosting. So, it is possible to connect a mouse, keyboard and screen and develop on the device. This is opposed to the more tedious process of cross-compiling, cross-deploying and debugging an exotic device.
1702845791×2
(Score: 2) by LoRdTAW on Friday September 05 2014, @11:26PM
Yes, it runs linux and has USB and HDMI. I own one.
The SoC is a Xilinx Zynq. That is a dual core ARM + FPGA on one die. The FPGA has dedicated IO ports to the ARM IO bus, not GPIO but straight to the bus allowing you to create your own high speed peripherals and use IRQ's or DMA. The Zynq has no video core but the system image comes with a pre installed bitfile for the FPGA that includes the Epiphany interface and a video core that outputs HDMI.
(Score: 2) by opinionated_science on Saturday September 06 2014, @01:18AM
Can we please have some benchmarks?
Linpack as a starter, then FFTW, gromacs etc...
This will give some talking points to discuss its place in the playing field...
(Score: 2) by opinionated_science on Saturday September 06 2014, @11:08AM
ok i read their website, and it rings a bell. I think I saw this talked about at a conference.
So they claim 90Flops/400W. Well that is a lot slower than a graphics card! However they also say it has FPGA logic on board too.
It will be interesting to see what the inter-process communication time is, because the amount of parallelism that can be extracted is dependent on this.
It might be useful as a system board with other denser compute elements. I seem to read this is what the "HP Moonshot" architecture is about...?
(Score: 2) by tonyPick on Saturday September 06 2014, @11:42AM
Benchmarking on the Adapteva site if fairly old, but references the EEMBC coremark.
Blog post is at http://www.adapteva.com/white-papers/more-evidence-that-the-epiphany-multicore-processor-is-a-proper-cpu/ [adapteva.com]
As I read it a single 16 core (Epiphany III) hits a coremark number just under an Intel i3 330M, and the 64 core is a bit below an Intel i7-3612QE.
That's pretty good result, considering the test isn't particularly well suited to the Epiphany (It's not floating point) and the differences in cost, form factor & power consumption. More up to date scores from someone who has the board would be nice though...
(Score: 2) by opinionated_science on Saturday September 06 2014, @04:07PM
Thank you for that!
Floating point is not *mandatory* for many algorithms, but certainly is convenient...
My desire for Linpack is simply that it gives a comparable number across a number of machine parameters - parallel efficiency being one of them.
(Score: 0) by Anonymous Coward on Saturday September 06 2014, @08:59AM
What the heck does that open mean?
(Score: 2) by c0lo on Saturday September 06 2014, @10:55AM
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 1) by Gravis on Sunday September 07 2014, @04:40PM
it's a computing platform that has a manycore coprocesssor and doesnt hide anything. no NDAs, no binary blobs, all GPL'd software(GCC)/firmware(HDL code for Zynq chip), the hardware platform itself is open hardware and the parallel processor (Epiphany III) it's very well documented in it's operation. the only thing that you cant have is designs to the Epiphany III itself.