Jon Smirl wrote:
On Fri, 18 Aug 2006 22:21:41 -0500, Patrick McNamara wrote:
howard parkin wrote:
I think the current goals of OGP and the needs of GPGPU are not really
compatible.
They want supercomputing power. Lots of compute power equals a big chip
(like the Clearspeed CSX600 - $750!). A big chip is expensive, which is
incompatible with the 'Open Graphics for the masses' goal of OGP.
I'm not so sure they are. Perhaps they are inconsistent with our current
approach. Perhaps that approach should be revisited.
The risk of a card for the masses is that the masses will actually buy
them, i.e. send in the cash, not just say they will send in the cash.
If a couple of high profile people start saying negative things they
can effectively kill your sales. What if AMD/ATI goes open spec? Your
main selling point disappears.
I pointed that out. The solution to this is (real) product differentiation:
X on the card
Ethernet card in a small box
Both together
Scalability (without having to buy more cards)
<etc.>
We can also, as TM pointed out, we can offer more and better support to
smaller customers. Open Source also means community support will be
available.
Going GPGPU looks lower risk to me. There are a limited number of well
financed players in GPGPU. Some of these groups even build their own
chips. The design can be pitched to each of these groups and you'll
get a definite answer. Either they will buy a couple thousand or
they'll tell you what is wrong with the design. These groups posses
high end engineering resources which will give valuable feedback.
Besides, you need to build the shaders sooner or later anyway. Nothing
is really lost.
See my other post: "A few words in favor of multiple chip architecture"
If you pitch all of the groups and no one is interested, then you have
your answer in a reasonable amount of time. This process is more
deterministic than "build it and pray that they will come".
Don't think that you need to beat NVidia/ATI in the first round. Just
build one really good shader, the GPGPU people know how to multiply
and figure out if it is feasible to squeeze 32 copies on a chip.
If we can offer a a fully programmable GLSL chip that does correct IEEE
math, we will have an advantage over nVidia/ATI.
Overview of the NV7800 architecture:
http://www.techreport.com/reviews/2005q2/geforce-7800gtx/index.x?pg=1
Start a thread in the forums at http://www.gpgpu.org/ That should let
you judge the interest level.
You also aren't limited to FPGAs. There is MOSIS.
http://www.mosis.org/
Good, such a service as I mentioned used to exist is still available.
With the educational discounts available it can
be quite reasonable for what you get in return. You have access to
cutting edge processes like IBM's 90nm low power process. You could
even get really lucky. IBM doesn't have a 3D chip design, they might
take a liking to yours and sponsor it.
Yes, if IBM would help, they have lots of money. :-D
There is also a lot to be said for software simulations before you
turn anything into a circuit. You could spend months coming up with a
cutting edge shader design. Do you want one large SIMD shader or lots
of little ones? Unified or separate fragment, vertex and geometry
shaders? Direct X 10.1 includes both float and integer support in the
shaders. Is the FPGA stage really necessary or is it possible to go
straight to silicon via MOSIS?
I would start with the fragment shader and have a standard CPU (possibly
with some extra hardware) feed them. Bottom up design. And
programmable fragment shaders are what GPGPU needs.
On another topic, it is possible to build a graphics card without VGA
support and a ROM. It's not like you are trying to run Windows on
these cards. Anyone interested enough to acquire one of these cards
can switch to a custom GRUB and kernel video driver that don't use
VGA. Most Macs don't have VGA support and no one ever notices when
they run Linux on them.
The problem is that PC MotherBoards expect VGA to boot and _configure_.
Not being able to configure your MotherBoard (and not being able to
see the error messages) is like a major issue. Now if they were all
well behaved and always used the Video BIOS calls, it would be no
problem -- the problem is that we are dreaming if we think that. The
easiest solution is to just use a VGA core. We could start with the one
from Open Cores:
http://www.opencores.org/projects.cgi/web/vga_lcd/overview
or we could use a commercial VGA core. Either way, we could concentrate
on the 3D stuff and not waste time reinventing the wheel.
--
JRT
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)
