Go Down

Topic: How about an Arduino style super computer the size of a credit card (Read 3 times) previous topic - next topic

westfw

I can't figure what to do with 84MHz of 32bit ALU on th Due, and you want me to contemplate a 66gips multicore?

MichaelMeissner

The problem with massively parallel computers is you need to have applications that have enormous data sets that you do all of the same processing on each subset of data.  You need to be able to break this data down into subparts so that each processing element can process its own subset of data without having to get bogged down because you are waiting for data from other nodes.

Sure, there are are various things that are massively parallel (rendering for instance), but I suspect unless you already have an application that is massively parallel, it isn't the chip for you.  When I was between jobs the last time about 4 years ago, Adapteva may have been one of the companies I was looking at, but it wasn't the right fit.

Ran Talbott

It doesn't have to be "enormous" volumes of data (except by 8-bit MCU standards): you could also do lots of processing on small-ish volumes.

Or do "soft peripherals" a la Ubicom (recently merged into Qualcomm): they used multiple cores to bit-bang interfaces like Ethernet (and even SATA, according to one story I read recently), instead of making many models with different specialized "hard" peripherals.

Let's say you want to do something like the Ardupilot:  dedicate a core to each servo and each input sensor,  and you never need to worry about your elevators wobbling if another part of the system is busy normalizing a temperature reading or passing data to/from a distant central PC.  And encrypting the data link so no one can hijack your UAV or steal the data you're collecting.

Similarly, for ground robots, you could create custom servo and motor speed controllers in software, with high-level commands like "move forward at 3.2 feet per second" or "wag the servo between 85 and 125 degrees with a period of 3.2 seconds".

64 cores is more than you'd need for the vast majority of hobby-level applications, but I could see even a relatively simple robot using 16 or more.

liudr

Kickstarter is a spam, this guy is a spam. Let's not waste brain power on this post.

Go Up