The Cell processor

[bascule]

IBM, Sony, and Toshiba will be publishing information on their upcoming Cell processor soon. Much of it, however, is already available here:

http://journal.pcvsconsole.com/?thread=9240

The fundamental idea is one of non-locality of computing resources. That is, you don't need to throw something away just because its processor is old, you can simply network it with newer devices in your home and the computing resources of those systems could be used in lieu of local ones. A Cell workstation could, for example, be linked to another Cell processor in a display, and 3D accelerated rendering could be performed entirely on processors within the display rather than those in the computer itself.

The entire architecture is pro-actively parallel in design, so I don't really know what programming language they intend to target these systems with...

[LosT]

This concept has always fascinated me. Creating building blocks, or units has always made sense- the question being at what point do you need to create 'specialized' cells- that is how fundamental you make your units and still accomplish your goal of being able to build something useful out of it. The BEAM robotics guy was pushing this concept with analog 'bug' robots quite some time ago-

I've always thought computers were just complicated LEGO ;)

LosT

[skroo]

The fundamental idea is one of non-locality of computing resources. That is, you don't need to throw something away just because its processor is old, you can simply network it with newer devices in your home and the computing resources of those systems could be used in lieu of local ones.

It's an interesting idea; from my reading of it they're essentially talking about dynamic clustering but in an ad-hoc setting. Obviously it's not a new idea, but seeing it implemented in hardware at this level of the market makes me wonder what Sony's plans are for all of this. Of course, much of what's in that article is assumption, so we should probably not lend too much credence to any ideas we come up with.

A Cell workstation could, for example, be linked to another Cell processor in a display, and 3D accelerated rendering could be performed entirely on processors within the display rather than those in the computer itself.

A la Plan9, but on a more organic scale.

The entire architecture is pro-actively parallel in design, so I don't really know what programming language they intend to target these systems with...

Dunno, but you can bet that the underlying OS will be heavy on thread/event scheduling.

[KeLviN]

makes you look forward to the play station 3....

i just hope its not another technology that works great in its limited distribution but then dies out before every one can get on the band wagon....

[TheCotMan]

IBM, Sony, and Toshiba will be publishing information on their upcoming Cell processor soon....

One of the biggest problems in such a technology is echoed in the history of hardware and system design.

1) Bus speed for inter-cell communication is likely to change with time whether the bus is physical or not. If backwards compatability is maintained, and protocols are changed, then each change will likely increase overhead and/or create latency by dropping communications to that of the weakest link, or in the least hinder performance with delays.
2) Assuming future reality shows Moore's Law will map complexity to proportional increase in speeds (meaning 2x complexity = 2x speed) then time works against cell processors arrays/networks over the long term.
3) Technology changes quickly. Consider SCSI1 through to SCSI3, all of the different connections, variations in distance/speed tradeoffs for LVD/Differential and legacy SCSI devices/drives and how using a 10 year old SCSI device on a new bus impacts performance on other new devices on the same bus. (Time issue again.)
4) True bus topology, ring, star, or "other" for cell-communication can be shown to have point of diminishing return for added cells, where cost of communication is more than processor savings. Even very intelligent schedulers can have problems working in complicated distributed networks. Different speed devices with different bus speeds? weighted directed graphs? This could lead to complicated routing programming if optimization for speed is desired.

I am impressed by the idea, but I wonder about its long term usefulness. There is promise in this idea over short time intervals (3 year EOL cycle for example.)

However it works out, cell processing it is a sexy idea: distributed local computing. :-)

[Added content:]
By "3 year cycle" I refer to the age difference of device in use at any time in a given system.

[kallahar]

Robotics.

Japan is betting its future on robotics, and one thing that robot AI's need is massive parallelism since so many things need to be interpreted at once. For example, if you have 500 sensors it makes a lot more sense to process them all at once and send an interpreted signal to the "brain" than to process them sequentially. I'm betting this is just the first gen of a long line of parallel-focused systems aimed at robots.

[converge]

.. makes me wonder what Sony's plans are for all of this..

.. a couple million or so playstations around the world acting as cells, interconnected for 'gaming' purposes would make for one shitload of a supercomputer. iirc very few games are written to actually tax the resources on a console, especially with modern technologies being thrown out in the market for gaming purposes. I smell spare cycles.

[kallahar]

Latency is the big problem, very few problems scale well to distributed systems. Mostly just the massive calculation problems like seti, folding, cracking, etc.

[Voltage Spike]

Latency is the big problem, very few problems scale well to distributed systems. Mostly just the massive calculation problems like seti, folding, cracking, etc.

That depends on who you talk to. Rodney Brooks (among others) might say this is a positive step in the direction of artificial intelligence. The latency of the human nervous system might be considered a problem in traditional computing, but most of us seem to do okay.

[bascule]

very few problems scale well to distributed systems

Actually an inordinate number of problems (anything involving nested loops, really) can benefit massively from parallelism, and the real limitation has typically been languages and development tools that can take advantage of this. While autoparallelizing C compilers tend to suck, there are some excellent autoparallelizing High Performance Fortran compilers.

A great deal of research in this area was done on the original Transputer, one of the first pro-actively parallel processor designs. A special parallel programming language was developed for it called Occam which resulted in easily programmed intrinsically multithreaded applications.

It's pretty much expected that Sony is going to need to develop an Occam-like language to allow the power of Cell to truly be leveraged as a language like C simply isn't up to snuff.

Cell is really what modern video games are begging for. Cell allows for multiple realtime tasks to be executed simultaneously, so you don't have to worry about attempting to effectively multitask them on a single processor while still concerning yourself with proper realtime execution. It's rumored that Sony will be pushing for deveopment on top of TRON-based cells, so Japan may soon be sporting a common realtime programming platform for virtually all of the programmers in the country.

[converge]

dood.. you just got me hard