512 bit memory bus is here :)

DaveBaumann said:
OICAspork said:
=D Does all this info you are sharing by any chance coincide with an immanent review/preview of the card?
Click to expand...

Nothing on the horizon just yet unfortunatly.
Click to expand...

I don't suppose random college student giving them a call on your behalf would help. :(
 
So, we here are speaking of aggregated 512bit bus? Each VPU has its own 256Bit bus, and together they can make an 'illusion' of 512bit bus? Or am I wrong?
 
Sunday said:
So, we here are speaking of aggregated 512bit bus? Each VPU has its own 256Bit bus, and together they can make an 'illusion' of 512bit bus? Or am I wrong?
Click to expand...

depends on how they handle the dual gpu interface I guess
 
The Realizm 100 and 200 have 32 GB/s.

The 3 chip Realizm 800 has AFAIK 16 GB/s for the VSU,
and 32 GB/s for each of the VPUs.

It should scale well as the rasterisation is tiled and there
appears to be a bus for exchanging tiles between the VPUs
to get them to the right DACs. At least that's what the diagram
appeared to show.

FWIW the Realizm 800 has a fair chance of beating the
Quadro FX 4000 on all the key CAD/CAM/CAE benchmarks,
in fact I think it'll be a while before Nvidia has a faster product.
Though Nvidia could be very agressive with pricing if they
wished.

I'd love to know how good the GLSL support is.
 
glw said:
The Realizm 100 and 200 have 32 GB/s.

The 3 chip Realizm 800 has AFAIK 16 GB/s for the VSU,
and 32 GB/s for each of the VPUs.
Click to expand...

Is it the VSU that has access to directburst then I take it?
 
remind them that Beyond3D is becomming more and more popular at CGTalk.com ;)

thanks to yours truely :D
 
Sage said:
remind them that Beyond3D is becomming more and more popular at CGTalk.com ;)

thanks to yours truely :D
Click to expand...
God why oh why have you foresaken us with Sage. LOL :p ( Just joking had to take a crack at you after that comment ).


I'm glad to see such a competitive product come out of 3dlabs I was a bit dissapointed with the specs of the previous generation. Then again I can't afford a X800 let alone a workstation board.
 
bloodbob said:
Surely very Large scenes ( distance wise ) tend to go crazy with single precision FP.
Click to expand...
If you have all your objects stored in their own local coordinate systems, then there should not be any problems with very large scenes and single precision.
 
Simon F said:
AFAIU it's got two rendering chips so I suppose each of those has a 256bit bus.
Click to expand...

That's my opinion too.

:me thinking: I don't uderstand why they disabled the VS units on the two 150mil tr chips. Waste of transistors. They could have designed a separate chip for the realizm800 realocating the unused transistors to embedded ram (i think P10 had, not sure tho).
 
Simon F said:
bloodbob said:
Surely very Large scenes ( distance wise ) tend to go crazy with single precision FP.
Click to expand...
If you have all your objects stored in their own local coordinate systems, then there should not be any problems with very large scenes and single precision.
Click to expand...
The objects are stored with FP32 precision anyway. But you have to transform all vertices into one common space eventually, so FP36 precision might give slightly better results.
 
Xmas said:
Simon F said:
bloodbob said:
Surely very Large scenes ( distance wise ) tend to go crazy with single precision FP.
Click to expand...
If you have all your objects stored in their own local coordinate systems, then there should not be any problems with very large scenes and single precision.
Click to expand...
The objects are stored with FP32 precision anyway. But you have to transform all vertices into one common space eventually, so FP36 precision might give slightly better results.
Click to expand...
I was not suggesting that each individual object use the full range of FP precision but, instead, only use a relatively small subset.

A base translation vector, for positioning each object, can be stored at double precision. Similarly, the camera location would also use doubles. When the SW sets up the local->camera transformation it can take the differences and convert these back to single precision. If an object is near the camera you still have "full" precision. If, OTOH, it's a long way off the precision drops but that doesn't matter.
 
You must log in or register to reply here.

Similar threads

S
Memory bandwidth vs memory amount *spin off*
2 3
Replies
42
Views
35K
Kb-Smoker
K
M
Will 512-bit bus + GDDR5 memory have a place in 2009 ?
2 3
Replies
42
Views
11K
CarstenS
CarstenS
M
  • Locked
mobility 9800 benchmarks!
Replies
4
Views
3K
Dave Baumann
Dave Baumann
Back
Top