Parallella: A Supercomputer For Everyone is Dying

It's closer to regular multi-core CPU computing than GPU computing. It's general purpose cores.

What sets it apart is that the cores are tiny, with little per-core memory (though all cores can transparently access each-others memory as well as main memory), and so the architecture is well suited for scaling up the number of cores with quite low power consumption.

So for problems that can be parallelized reasonably well, but with more complex data dependencies than what a GPU is good for, this might be a good fit.

I'd put it somewhere in the middle between GPU's (for embarrassingly parallel tasks) and general purpose CPU's with high throughput per core.

Also, this looks like it'd be possible to fit in the power envelope of really small embedded systems, like e.g. cellphones and tablets....

Before more developers have these systems, it'll be hard to say how useful they'll be, but the architecture looks exciting.

That's why I supported it - I really want to see how this type of architecture can be exploited, and whether or not it'll prove to be cost effective and/or simpler to work with than GPU's for the right type of problems.

Typical multicore CPUs don't have nearly as many cores as Parallella. Also, from what I can from www.apteva.com/introduction, the power consumption is much lower and the interconnect is different. In Parallella, cores are laid out in a grid and cores can only talk directly to their neighbors.

Vs Xeon Phi: Cost, complexity, power. Look at pictures of the Xeon Phi cards. They're covered in heat sink, and with a fan. For comparison, the Epiphany chips are a single tiny die with no cooling. But of course the per-core performance is not likely to be anywhere near Xeon Phi either.

I'd consider Epiphany the simple, "slow" (per core), low power solution, with Tilera somewhere in the middle, and Xeon Phi at the other extreme (complex, fast per core, high power usage).

That said, this is speculation based on reading articles - I've not had my hand on any of the three. Yet :)

At some point in the next 24 hours, the backers who signed up for one unit will need to ask themselves if they would rather have two units or zero.

I can't speak for the other 1800 people in my bin, but I just decided on two.

I've done my master thesis on GPGPU, so maybe I can help out a bit. I'm not yet too familiar with Epiphany's design however. From what I could grasp what sets them apart the most is a different memory architecture compared to multicore CPUs, where the individual cores seem to be optimized for accessing adjacent memory locations as well as the locations of the direct neighbors. This is one point where the architecture seems to be similar to GPUs, although GPUs have a very different memory architecture again - for the programmer it might look similar however, especially when using OpenCL.

The main point where Epiphany is diverging from GPUs is that the individual cores are complete RISC environments. This could mainly be a big plus when it comes to branching and subprocedure calls (although NVIDIA is catching up on the later point with Kepler 2). On GPUs the kernel subprocedures currently all need to be inlined and branches mean that the cores that aren't executing the current branch are just sleeping - Epiphany cores seem to be more independent in that regard. I still expect an efficient programming model to be along the same lines as CUDA/OpenCL for epiphany however - which is a good thing btw., this model has been very successful in the high performance community and it's actually quite easy to understand - much easier than cache optimizing for CPU for example.

If we compare epiphany to CPU what's mainly missing is the CPU's cache architecture, hyperthreading, long pipelines per core, SSE on each core, possibly out-of-order and intricate branch prediction (not sure on those last ones). The missing caches might be a bit of a problem. The memory bandwidth they specify seems pretty good to me, but from personal experience I'd add another 20-30% to the achievable bandwidth if you have a good cache (which GPU has since Fermi for example). The other simplifications I actually like a lot - to me it makes much more sense to have a massive parallel system where you can just specify everything as scalar instead of doing all the SSE and hyperthreading hoops like on CPUs - optimizing for CPU is quite a pain compared to those new models.