Parallella Community

I am a newbie for parallel computing and wondering what is the difference between parallella and GPU?

GPUs can only be used for data parallelism, with no communication between cores. The Epiphany cores are full-fledged CPUs that support data, task, and other types of parallelism, as well as communicate with each other to perform more complex computations. As an analogy, GPUs could be seen as factory workers performing one task over and over, while Parallella is a group of artists working together on a project.

Surely the obvious difference is power. The Epiphany IV draws about 2W for 100GFlops. I think you can get the same performance out of a top end Nvidia GPU, but drawing more like 50W. So you can use Epiphany in embedded applications.

As a comparison, Titan, currently the worlds fastest supercomputer at nearly 18 PetaFlops draws approximately 8MW. On a naive calculation, if it were built using 200,000 Epiphany IV chips it would draw just 400 kW. Or to put it another way, a more than $7M annual electricity bill reduction.

Happy New Year,


Jeremy

so having a system with 800 Epiphany IV cores gets us well and truly into the Top 500 cool. $100,000 for a in the Top 500 supercomputers. assuming $100~ will cover hardware for each core, Ethernet and routers. probably will not unless theirs multiple cores on each PCB, like 10 Epiphany IV cores per PCB with a host Arm CPU. so that would be 80 PCBs
In the end it will still be under 2KW power draw. Just over 2 gaming PC's.

Correct me if i have made a careless mistake.

really? terabytes per core? would be interesting then to have an arm-proceessor that uses 64-bit and has those terabytes available, so the coprocessor's cores could indirectly address that additional memory even though being still 32 bit themselves. quite a challange to implement the dma-transfer though, and price would be way beyond 100$ for the memory alone...

anyway, parallella is made by a person who worked at signal-processing. of course that will be the only thing it's good at! however, signal-processing is what our brain is doing all the time, with very little actual storage per neuron. so maybe those 99% HPC problems came into existence only because there was no actual low-mem signal-processing supercomputer? that's always the problem with using statistics for planning economy, you never know if the lack of some product didn't bias the statistics against that product's introduction! remember what someone once said about the invention of computers, that only a handful of them worldwide would be sufficient. now everybody has one...

I see an advantage to the Epiphany architecture as the OS runs on the ARM host CPUs (external for now) you can just have pairs of dedicated cores in the Epiphany to do High side and Low side Data.

So Two 32 bit cores can do 64 bit data operation or Four cores can do 128 bit data operations, as the cores are not burden with the OS running in the background you can tell them to do anything even with a closed OS on the ARM host CPU.

This is something i do not believe other high end supercomputers can take advantage of unless there using a GPU. but obvious sacrifices are your dividing the effective cores in the Epiphany.

News about 64-bit ARM:

How to cheaply acquire TBs of RAM:

Another major advantage is scalability: just compare Adapteva's roadmap to Intel's and I'm sure you'll like a 10x increase in processing power over 10%. Keep in mind that the Epiphany is not so much a GPU, but rather more like a fledgling CPU, if you will.