Parallella Community

From what I understand there is the 16 core chip and board, then there is the 64 core chip that uses the same Zynq board.. both will be available in 2014?

Will there be a Parallella cluster kit of four 64 core boards for about 600 dollars? Time schedule? For a total of 256 cores? Is that correct?

What interests me are AI applications like IBM's Watson. Watson was claimed to have 2,880 POWER7 processor cores and had 16 terabytes of RAM creating performance estimate of 80 Tflops.

45 boards with a single 64 core chip equals 2,880 cores, but would they be the comparable to the 2,880 power 7 ibm cores? 45 boards at 100 dollars each would be a cost of about 4,500 to 5,000. In 2011 it toook 3 million for 2,880 power 7 cores. I do not know how or if the Parallella chips could or would access 16 TB of memory. Is it possible for this type of application with Parallella, Provided you used OSS for the AI, could you get the similar 2 to 3 second response time per question? Or are Parallella chips not suited for that type of application? (hypothetically speaking, since it has not been done yet.)

16 TB of memory would be 500 32GB sticks of ram!


I do not know if you could get the same performance speed with Parallella that IBM did with watson, but big data and AI are very interesting. It would be wonderful to see the type of projects and innovation that could a occur, especially if the hardware that allowed semantic AI type projects like watson could be done inexpensively.

How will Adapteva's round of funding change priorities and pricing, if at all?

Would you consider Parallella a disruptive technology, or is it too early in the development cycle?

Don't what to shatter you illusions, but each Power7 core is 99.36 GFlops, each chip has 8 cores so 794 GFlops. One 64 core, parallella is 102G Flops, about the same as one Power7 core and only has 2Meg of Ram. My understanding of what the Parallella board is aim is an affordable board that hopefully will allow people to develop parallel processing techniques without having to invest in large machines.

The 64-core Epiphany chip delivers a total of 102 GFLOPS (1.6 GFLOPS per core):

http://www.adapteva.com/epiphanyiv/

I think I understand it now, if you had the proposed 1K core Parallella chip/board, you would need 80 of them to get 80 Tflops.

To get watson's 16 TB of ram memory each 1k board would have to have 32GB memory x 6. In today's dollars that would be too expensive.

32GB times 500 ram memory sticks equals 300,000 dollars at current pricing.

It will be interesting what kind of semantic web/AI applications you will be able to create with this platform.

If we would just get the 16K core chip version with the proposed 20 Tflop performance, 4 chips and you would have 80Tflops.

We just have to wait till 2022...8 years is too long of a wait.

Except that each Epiphany core only has 32K memory (Yes that is K not M) so 80 * 1024 cores * 32K only amounts by my calculations to just over 2.6 Gigabytes of Memory. Not in the same league as 16TB ram unless their is a dramatic change of Parallella architecture.