As a scientist used to running simulations on parallell computers, my immediate response is "show me one use case where this is faster than four sandy bridge core i7 desktops connected with Gb ethernet using OpenMX to get low latency". Even though you can boast a decent core count, those cores are connected through a really crappy high latency network (the internet).
Basically, there is no scientist in the world today who doesn't have the funding needed to scrape together four used core i7s and network them. Unless you can significantly beat that performance level, you're not "democratizing access to supercomputing for the scientists that need it" in any meaningful way. If you are significantly faster, even for just one particular application, that should be up front and center on you FAQ.
Are saying that folding@home, BOINC and all the rest is all useless and that all scientists need in to do research is 4 old i7 desktops? As far as the benchmark is concerned, check out this presentation from Ericsson on page 9 concerning FFTs on Xeon and on Epiphany. http://wiki.portal.chalmers.se/cse/uploads/FP/EngdalSlides.p...
No, the @home stuff is leveraging thousands and millions of computers. You have 18. From the slides you linked, the Epiphany board is 27x slower than a Xeon, so your distributed supercomputer is about as fast as a single desktop with a core i7, assuming you have perfect scaling.
The slide stated that one Epiphany core is 27 times slower than one Xeon core. clearly the goal here is to ramp..the 324 cores all came online today. If we don't get to >10,000 cores, then the effort has failed.
Will you send out any test load before May 30? Would be better to see those cores to work if they are online, otherwise it feels kinda wasted. Wouldn't be surprised if people go offline after a while if there's no work done.
So, I'm confused. That slide says 64 cores, but the Parallella board website says 18. Which is it? You really need a proper benchmark.
Another question: how are you planning to implement fault tolerance? If you're running across hundreds of nodes via the internet, the probability of one failing while my job is running is high. Are you going to run a fault-tolerant scheduler?
And: how are you going to do file I/O? Does the user have to run the master MPI process on his/her own machine and do I/O there?
I think the 18 is counting 16 Epiphany cores plus 2 cores on the Zynq. The 64 is referring to a 64 core Epiphany chip that has only been made in prototype quantities, so far as I know.
The efficiency of these board (Tflops per watt) is pretty amazing, when I first saw this board I imagined it would be used as powerful embedded servers in situations where battery power is required. However, it seems they have a much broader use cases as well.
What application will be deployed on this for the live test on May 30th?
When will we get a demonstration with a real interconnect? If the FLOPS/Watt is as good as claimed, why is nobody packaging this as an actual supercomputer?
Epiphany is being considered as the technology that comes "after the next". CPU is the default, then GPU, then "the others". Software is king and shared resource supercomputers have an incredibly long list of legacy software requirements. This is why we are focusing this effort on an application that is pretty dynamic at the moment, machine learning.
because small, slow computers are easy to make power-efficient. the truth is that the processor isn't that interesting, if you're taking a lots-of-wimpy-nodes approach (like BGQ).
Do you have some ideas/opinions who the industry will contain power consumption with big cores? Current super computers running big brawny cores and GPU accelerators are running at 5 GFLOPS/W at best. (consensus that we need to get to 50 GFLOPS/W).
