Re: Moore versus TPL: The bomb has burst

Albert van der Horst <albert@xxxxxxxxxxxxxxxxxx> writes:
Apparently the business "partners" of Moore have collected (and
squandered) some $300 million in royalties for the MMP portfolio.

Well THAT is interesting. I wonder how much of that was related to the
Forth array stuff, rather than (e.g.) the a-d converters.

The pdf files on Greenarrays website is an interesting read,
especially the perspective of picojoules per operations.

I see the page

http://greenarrays.com/home/documents/greg/WP003-100617-msp430.pdf

compares a GA array processor with a low powered general purpose TI
microcontroller. It says the GA chip does a 17x17 bit fractional
multiply with about 450 pJ of energy.

Let's try the same calculation for a current semi-mass-market graphics
coprocessor (NVidia GTX 480M Mobile Fermi):

http://www.anandtech.com/show/3740/nvidia-announces-gtx-480m

100 watts power consumption, 598 gigaflops
= (1e2 joules/sec) / (5.98e11 ops/sec)
= 1.72e-10 joules/op = 172 pJ / op.

So the Nvidia chip is already using less than half the energy per
operation of the GA chip, assuming (as is traditional in the DSP world)
that these flops include multiplications. Also the NVidia is doing
32-bit floating point arithmetic, so quite a lot more work per operation
than the 17-bit GA arithmetic.

Of course the 100 watt Nvidia part is not directly comparable to the
tiny GA4, but the most interesting GA product has always seemed to me to
be the big GA144 arrays. They have announced a little board with nine
GA144's on it and I thought of some possible uses for it, but in each
case, conventional DSP's seemed to make more sense. I wonder why they
don't compare the GA with a small DSP, since that's the obvious area
where you want a lot of MIPS and not a lot of memory. The GA's don't
have enough memory or i/o bandwidth for the compute-intensive
non-numeric tasks I can easily think of, though maybe I'm missing some.

I never had the impression that there was any magical process technology
in the GA chips, just clever and unique (Forth-oriented) CPU
architecture. But basic ALU operations are very well studied by now.
So I don't see how a part made on a 180 nm process can possibly beat a
part made on a 45nm or even 32nm process no matter how clever it is.
.

Relevant Pages

  • M1330 laptops nVidia chip overheating?
    ... My XPS M1330 laptop seems to have heat problems ... with its nVidia graphic chip. ... But then the laptop worked fine, ...
    (alt.sys.pc-clone.dell)
  • Re: Fedora 13 on Dell E6410 screen resolution problem
    ... nvidia driver does not support this chip. ... on the list of supported chips at the Nvidia web site. ... Has anyone actually gotten Fedora 13 with X to run properly on one of ...
    (Fedora)
  • Re: Nvida card
    ... Add to this the fact that nVidia will sell its chips and reference hardware design to anybody at all, and you end up with all kinds of nVidia graphics cards made to various levels of quality by manufacturers of good and ill repute. ... If the card is used in a dusty dirty environment with lots of animal hair or feathers, the fans clog up, stop running, and the graphics chip burns up. ...
    (alt.sys.pc-clone.dell)
  • Problem with re0
    ... I have an Asus M2NPV-VM motherboard with integrated Nvidia MCP51 Gigabit ... TP-Link TG-3468 PCIe network card which is using Realtek 8111 chip. ... re0: Using 1 MSI messages ...
    (freebsd-net)
  • Re: x,y addressable LED matrix
    ... Does anyone know of a suitable array? ... If a number of these control circuits could be put ... >onto a chip, one could avoid the need to have a chip for each LED. ... A common scheme is to arrange the LEDs into a Cartesian 64X64 array ...
    (sci.electronics.design)