What is the reason Parallella isn't blazing a trail ?

[theover]

It could be timing, not making the promised spec, it could be the USB trouble, the distribution and price, etc etc, but what is really holding the Parallella back from becoming the "new supercomputer" ?

Theo V.

[nickoppen]

I disagree. The parallella is blazing a trail. This is what trail blazing always looks like.

Whenever you leave the mainstream, there might be a bit of publicity but the mainstream does not care - it just goes on doing what it does and with the weight of numbers, it will still dominate the chatter (press, social media etc etc). From the point of view of the new venture, Parallella in this case, there is always a mountain of problems that you have to solve that might not have been obvious when you set out. A good example is the PAL project. Solving all of these problems takes time even for companies with large resources.

Once a significant number of these problems have been solved, you then have to find a niche where you can start to generate real commercial returns for customers. By that I mean solve real world problems in a way that is significantly better value than any alternative so that the cost of transitioning to the new platform is worthwhile. (Selling boards to enthusiasts and research labs is only a short term solution for a company. That market will dry up pretty soon.)

Then, once the niche has been established and the product is mature and well understood, there is a chance that there might be wider adoption. I think that the Parallella and the e-16 is a great first step and that Adapteva's strategy is spot on for this phase in the evolution. Yes, the technical issues need to be solved, the price/performance needs to be right but a lot more needs to happen before the Parallella is the supercomputer for everyone.

[piotr5]

apart from lack of popularity, imho only lack of social-networking is holding parallella back. amd had this idea for a hsa library, parallella has its pal library. and there is also the odroid-xu3 which uses 4x2Ghz A15 cores from samsung and 4x1.4 cortex-a7 with about the same price as parallella but 2GB ram, and attempts to combine this heterogenous system through a linux-scheduler. why can't these 3 players work together on a single software solution? somehow c++ got standardized now, why can't there be a comittee for standardizing a parallell computation solution? the answer is simple: the people involved are all much too introvert to initialize such a global communication process. you know, the trick is to not just walk up to amd and say you want to cooperate, instead send a whole crowd of developers who agreed to work on some standard. here it was said, only 100 functions are needed for pal, but every developer could tell you that for practical use you actually must adapt such a library to the software using it, this is a complicated process of software design. to do that you actually must meet with developers and listen to them describing their projects, and preferably there should be a bunch of people collecting information that way. when did this happen? this kind of research doesn't require popularity, it requires to meet people and to have friends who could be convinced to help in such an attempt. in other words, adapteva needs a better PR department, one which values relations to developers over general popularity. not everything in the world can be bought with money...

[Gravis]

[piotr5]

[Gravis]

[mhonman]

[piotr5]

my biggest wish for Epiphany V would be that adapteva cheats: a chip with 16x16 ecores, but this time double the assembler-command-set to offer 16-integer vector computation additionally, and of course bandwidth of eLink would need to rise accordingly. this would create the illusion epiphany would have 4k simd cores. (you know a graphics-card which has that?) that would attract simd-programmers to learn mimd programming. (16-integer vector is really tiny, in comparison to the 256 actual mimd cores.) and it would directly provoke amd which offers only 4+32x16 simd in its apu. did you read what it takes to create a motherboard? basically it's trial and err. the reason for that is insanely high data-transfer rates. so I agree, integrating some system-stuff would be nice. again amd is starting out with some 3d memory without realizing how valuable that would be in mimd parallell programming. epiphany clearly could stir up the market there.

but as I know adapteva, they refuse to tread in the footsteps of giants like amd. still, some kind of 1k cores epiphany should come out soon, they'd better work on it already...

to pull out survival of the fittest again: adapteva is a hardware company. if you see a lack of software for their products, that's your own fault for not doing that programming yourself. intel didn't do it right, they created a heuristic compiler which has to run the software for fine-tuning. a real programmer could do that better! thanks for the pointer at xmos btw...

[sebraa]

I guess then it's time to get your hands dirty, and show us what you've done...

[mhonman]