[extropy-chat] Greetings

[Eugen Leitl]

On Fri, Dec 09, 2005 at 09:39:23AM -0700, Derek Zahn wrote:

> ? I am under the impression that the SPEs have very nice floating point
> units; the cell processor has a theoretical rating of over 200 single
> precision GFLOPS.  Of course, that's absolute best case, but it's

The float performance is very good indeed -- but floats are overkill
for AI requirements. Both the number of bits (state) and the dynamic 
range especially. Biology works with few bits resolution quite successfully.
Short integers are both more compact (less memory bandwidth) and can
profit from SIMD-in-register parallelism.

> amazing to me nonetheless.  Additionally, the PS3 GPU should be able to
> do another couple hundred GFLOPS in the pixel shaders (also best case,
> etc etc).

Yes, and GPUs have very good memory bandwidths in the bargain
(XBox360 has 250 GBytes/s bandwidth, albeit only a small amount of it).
GDDR4 trends towards similiar bandwidths, however. 

Despite vortex shader languages they're not nearly as comfortably 
to program as with all-purpose CPUs.
 
> Yup, that's my hope.  I am a bit skeptical that Sony is going to be
> eager to sell linux boxes at a loss so I'll believe it when I see it.

The first two years will be lossy, and the last time they shipped a PS2
Linux kit with a hard drive and NIC which pushed it well into profit range.
Number crunch is a tiny niche, so it might not cut into sales that much
even if they ship the kit soonish.

> Until then I don't really want to spend time playing with a simulator.
> If that doesn't work out, the gaming industry has produced similar
> astonishing computation densities in the GPUs of PC-based graphics
> cards.  The Cell looks really cool though and I'd rather use that.

Yes, it is rather interesting.

> Other non-PS3 Cell systems (like blade servers) might be barely within
> the budget of a serious hobbyist.  I'm content to wait a while and see
> what happens.

Ditto here. I'm sticking with AMD64 for time being.
 
> > So what are you going to do with it?
>  
> Play.  Dabble.  Muse.  Those outrageous GFLOP numbers would never
> translate to LINPACK results primarily because of bandwidth constraints
> but I'm curious about whether AI "substrates" could be designed
> specifically to make best use of these architectures, perhaps modeled
> roughly on the cortex.  I think it will be fun to look at.

It is possible to build a 3d array of mostly-local-connectivity neurons
which would be well-behaved in terms of the memory access pattern. Cubes
of those could exchange adjacent layers via MPI, and map well to GBit
Ethernet connectivity. Relatively addressed spikes (offset in the array and
a time stamp when issued) can be stacked tightly in a packet.

What's missing is a good morphogenetic code and a selection process
to mutate a few 100 neuron classes with interesting properties.
 
> In the meantime I'll probably spend the next year or so building a
> legged robot with video cameras and touch sensors and so on; I have
> built a number of "robots" before, with up to 25 degrees of freedom, so
> it won't be too difficult and this will give some focus for specific
> things to do with the CPUs.

Good luck, sounds like an exciting project.

-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a> http://leitl.org
______________________________________________________________
ICBM: 48.07100, 11.36820            http://www.ativel.com
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20051209/dbe1c78c/attachment.bin>