On Sat, Dec 17, 2011 at 12:04:48AM -0300, Jecel Assumpcao Jr. wrote:
> Eugen Leitl wrote:
>
> > [human limits - growing code]
>
> Perhaps the copycat work Doug Hofstadter, did is a step in this
> direction?
>
> http://en.wikipedia.org/wiki/Copycat_%28software%29
Thanks for the pointer, I haven't heard of that.
It would be interesting to see how Koza's GP would
work on realtime selection of very large agent
populations.
> > FPGAs suffer the problem of lack of embedded memory. Consider
> > GPGPU with quarter of TByte/s bandwidth across 2-3 GByte grains.
> > You just can't compete with economies of scale which allows you
> > hundreds to thousands of meshing such with InfiniBand.
>
> You can have the same external memories with an FPGA as with a GPU. I
> don't think that is an important difference between them.
The commodity/off the shelf (COTS) aspect is important, though.
E.g. a common graphics card buys you for 300 EUR some
Up to 880MHz Engine Clock
2GB GDDR5 Memory
1375MHz Memory Clock (5.5 Gbps GDDR5)
176 GB/s memory bandwidth (maximum)
2.7 TFLOPs Single Precision compute power
683 GFLOPs Double Precision compute power
TeraScale 3 Unified Processing Architecture
1536 Stream Processors
96 Texture Units
128 Z/Stencil ROP Units
32 Color ROP Units
Dual geometry and dual rendering engines
High Speed 256-bit GDDR5 memory interface
PCI Express 2.1 x16 bus interface
There's no problem putting 2-4 of these in a
single computer, shelve hundreds to thousands
of these and link them up with low-budget
InfiniBand. The bang for the buck is really hard
to beat with custom architectures.
> > [large clusters with COTS]
>
> Ok, but the networks used in clusters are a bit slow for the experiments
> I want to do.
Interesting. End to end latency in InfiniBand is almost exactly
a microsecond, 10-40 GBit/s thoughput allows you mesh topologies
and all kind of funky message types. You're obviously working
directly at FPGA level and need ns latency. What kind of work do you do?
_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc
