Wish List: 4GHz Propeller

Kirk Fraser

I watched Watson winning Jeopardy on PBS http://www.pbs.org/wgbh/nova/tech/smartest-machine-on-earth.html see also http://www-03.ibm.com/innovation/us/watson. To respond as good as or ahead of human contestants required many parallel processors because a regular computer took half an hour to give a correct response. So we students, hobbiests, and garage inventors need a cheap, easy to network, high power parallel processor. Other 4GHz processors exist so it is possible.

prof_braino

If you want parallel hardware, the issue is software to implement processing the algorithm in a distributed environment. We've already got a high speed link and protocol to make several props appear as one big prop with lots of cores. This is the easy part. The interesting part is next, where we try to figure out how to make all these cores work together on the same problem in an efficient manner. The resulting software would apply to prop1 or prop2, but it might take a while. But even if you did have a 4Ghz processor tomorrow, we'd still be lacking this software.

rod1963

Nvidia's CUDA Processing technology is pretty affordable. But it's for savvy coders who really have a specific app in mind that can benefit from the parallel tech. Dabblers need not apply. Just go check out Nvidia's offerings and see for yourself.

Kirk Fraser

Parallel OS is a solved problem, it's a matter of the application details. A Propeller already has code for dealing with cogs. When connecting N-Props you could treat them like more cogs on the rotor or you can try to create a more application specific division of labor. A good place to start would be Google's cloud code system, which handles their server blades. They have a way to learn it even though their own clouds are proprietary. Although I haven't dug deeply into Nvidia's stuff, I get the impression its very numeric oriented for graphics, operations research, and similar scientific applications. I'm less certain that approach would help provide hobby level people a version of Watson or better. For that each processor needs to access a lot of text from a mass storage device.

rod1963

If hobbyists are interested in doing some sort of home brew Watson, then Nvidia's CUDA could provide the computational assets needed. The software and mind share is certainly there for it.

The Prop isn't even a contender.

Heater.

Kirk Fraser,

Why is it that when people see "Propeller, 32-bit, 8 cogs" they start dreaming
of building highly parallel super computers out of Propellers? Whilst a faster
Prop would always be welcome it is totally the wrong architecture for such jobs
as Watson. It is an MCU first and foremost.

For tasks such as Watson you need a lot of speed and a parallelism. You are
going to need access to a huge database of some kind. To handle all that data
you will need some gigbytes of RAM on each CPU. That dictates a bus interface
to RAM. To allow the processors to work with the relatively slow off chip RAM
you will need cache memory on chip. Caching immediately destroys the timing
determinism of the processor which is one of the propellers key features. To
keep the processor humming you will want a deep instruction pipeline and
branch prediction and all those other tricks modern processors employ. Again
destroying determinism. Whilst we are at it there is probably no point in
having all those general purpose I/O pins or timers or video generators on a
chip built for Watson type problems.

In short, speeding up the Prop is not enough for what you want. Bending the
Prop to get you there destroys it's "Prop nature" if I can call it that. The
resulting device would be just another Intel/AMD, or ARM or IBM Power chip.
Which I'm sure is not a market segment Parallax want's to chase as it is well
covered already.

As a practical matter a 4GHz Prop is not going to happen until chip fabs are
offering such technology to companies like Parallax. Parallax has to work with
what is available and economical. Working with a 4GHz chip might be a bit
tough for us hobby constructors.

As an aside, I was just checking the specs of a new Power Architecture chip
from IBM. As might be used for Watson. Wonderful device, 16 cores each running
16 hardware scheduled threads. A lot of GHz. The cores runs at 1 volt and suck
64 amps. Do you really want a Propeller like that?

SRLM

Kirk Fraser wrote: »

Parallel OS is a solved problem, it's a matter of the application details. A Propeller already has code for dealing with cogs. When connecting N-Props you could treat them like more cogs on the rotor or you can try to create a more application specific division of labor. A good place to start would be Google's cloud code system, which handles their server blades. They have a way to learn it even though their own clouds are proprietary. Although I haven't dug deeply into Nvidia's stuff, I get the impression its very numeric oriented for graphics, operations research, and similar scientific applications. I'm less certain that approach would help provide hobby level people a version of Watson or better. For that each processor needs to access a lot of text from a mass storage device.

You're missing the point here. It's not the OS that is the issue that Braino was talking about, it's the specific program that needs to be parallel.

Cluso99

Heater. wrote: »

As an aside, I was just checking the specs of a new Power Architecture chip
from IBM. As might be used for Watson. Wonderful device, 16 cores each running
16 hardware scheduled threads. A lot of GHz. The cores runs at 1 volt and suck
64 amps. Do you really want a Propeller like that?

Yes please as long as it has the propeller architecture and costs <$20, but hopefully I could slow it to 1/4 speed and reduce current <<8A.

PS Ask a silly question and........