Yummy new CPU/GPU combo
Carsten Haitzler (The Rasterman)
raster at openmoko.org
Thu Jun 5 11:29:10 CEST 2008
On Thu, 5 Jun 2008 09:33:58 +0200 Tom Cooksey <thomas.cooksey at trolltech.com>
> On Wednesday 04 June 2008 21:56:56 Andy Green wrote:
> > | IMO, OpenMoko's choice of using the the glamo was a big mistake
> > (Connecting it to a
> > | shared, 4-bit bus was probably the _biggest_ mistake).
> > Huh what? It's a 16-bit memory bus, maybe you mean 2^4 ;-) When I
> > actually use the thing I don't notice much sluggishness. Your best bet
> > is to eyeball one, I think you'll find it isn't the issue you think it is.
> I thought it was connected to the (4-wire) SDIO bus?
it's on the memory bus, not sdio. a gfx chip on sdio would indeed be just about
the most stupid idea i have ever heard of and anyone doing that should be
summarily executed with a blunt spoon! :)
> If it's connected to the address/data bus, why is it so slow and why can't
> DMA be used to copy data to it?
dma can. it just is even slower than using the cpu. also the entire memory bus
is locked up when doing a copy with cpu OR dma, so even if dma were the same
speed as doing a copy with the cpu, we'd still have the cpu locked out of the
memory bus while waiting for dma - so no gain anyway (as in all practice you
will need the memory bus :)). actual tests found dma to be LESS than 1/2 the
speed of doing a copy with the cpu. this is just a limitation of the SOC, and
the glamo is just well... not fast at accepting read/write requests - the
bandwidth is limited to about 7mb/sec for writes ( can't remember what reads
are - but i think its symmetric).
so a big performance bottleneck we have is transfers to/from glamo - it really
hobbles performance as it is a tiny straw we send data through. not to mention
sd-card IO data also shares this glamo bus... :) so don't go doing any heavy IO
on the SD card AND gfx... at the same time! :)
Carsten Haitzler (The Rasterman) <raster at openmoko.org>
More information about the community