On 4/22/07, James Richard Tyrer <[EMAIL PROTECTED]> wrote:
Rogelio Serrano wrote:
> is it possible to create a gpu that can use up to 1 gb of system memory?
Actually, the question is about the system bus. AFAIK, all flavors of
PCI allow bus masters and would, therefore, allow a GPU on a card to
access all system memory.
The problem is the question of the speed at which you could do it. The
system bus becomes a bottle neck. Fast (66MHz) and wide (64 bit) PCI is
going to be less than 533 MB/s -- how much less depends on how many
addresses are issued.
If you are accessing textures as read only, you could use a cache on the
graphics card to speed things up. Since PCI flavors do not provide for
cache coherency, the cache would have to be cleared by the video driver
when a new texture was loaded into main memory.
AGP specifically addressed this issue and does come in 8x and wide.
IIUC, PCIe would be faster if you used 8x.
I think that HTX is the real answer to this issue.
yes and no. hypertransport is a point to point link in contrast to pci
which is a shared bus. but htx is just a link. even you can get to the
memory very rapidly you might have to wait until somebody else gives
up the memory. for a start yes but we need multiport memory to speed
it up.
the main point is i want to avoid copying window data from system
memory to the graphics card 30 times a second. so the graphics card
can render the data. i want the graphics card to read the data in
system memory one by one and immediately start emitting a video
signal. i dont want the graphics processor to wait when the copy is
done and then start reading the data in its local memory and then
start emitting video signals. imagine decoding video to a buffer in
system memory and then dma that to the graphics card 30 times a
second.
i dont want 512m mb on the graphics and another 512 in system memory.
when im not playing a game or watching a movie majority of the 512 mb
is unused. and when the graphics card needs more memory it can use
some of the other 512 mb. well we can sort switch memory around but
its not efficient and hard to get right and optimise.
in an asic based system maybe but not in an fpga based system. it will
be some tens of years before we will see any of this in an asic. i
just want an open computer system. i want access to all the functions
of the entire chipset and its subsystems. i want complete control over
the integration of the hardware and the software. i want a system that
is as integrated as the macs that steve wozniak put together.
--
JRT
--
the thing i like with my linux pc is that i can sum up my complaints in 5 items
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)
