[LAD] TI throws DSPs at supercomputers

[Charles Henry]

On Fri, Dec 9, 2011 at 7:08 PM, Jens M Andreasen
<jens.andreasen at comhem.se> wrote:
> From El Reg:
>
> - Nvidia had better watch out. Texas Instruments is not only its rival
> when it comes to making ARM processors that might end up in servers
> someday, but it is also repositioning its digital signal processors so
> they can be used as math coprocessors for standard x86 CPUs – and
> perhaps ARM processors one day.
>
> ... TMS320C66x family of DSPs needs a much easier nickname if it is to
> become cool and talked about ...
>
>
> C66x delivers 160 gigaflops of single-precision floating point ...
> 8MB L2 cache ... 12.8GB/sec of bandwidth ...
> only consumes 10 watts ...
>
> www.theregister.co.uk/2011/11/28/ti_dsp_supercomputer/print.html

Those are nice specs, especially the power and bandwidth--however, a
Nvidia card that does 160 GFlops is probably a *lot* cheaper (just
based on costs I've seen with TI DSP processors).  A card capable of
~500 GFlops only cost me ~$150, last year (mid-range card, previously
released series).  ECC memory adds a ton of cost--you have to buy the
Tesla series to get cards with ECC memory which is useful for
scientific calculations.  The power cost per flop is important in the
cold room, but to a machine for personal use, the power consumption
really doesn't matter.

I would really hope to see more competition for the GPGPU market soon.
 Most people see that Nvidia got out ahead of the curve.  I'd fully
expect other companies (I think, Intel, AMD, IBM, and TI) to have
other similar OpenCL compatible processors that can compete.  But not
today.

Chuck