ATI Announces HyperMemory

[Siskods9]

http://www.avault.com/news/displaynews.asp?story=9172004-92222

[retsam]

i dont get this, are they saying that they have pateneted an api that will help aps with faster access to and from the cpu and gpu or is this hardware that will make all the decisions... sort of like a switch of some sort .... mmmmm
/start_yoda_voice many questions there are many questions /end_yoda_voice

[Cruel_Logic]

i think this gain will be univeral when switching over to pci express... I'm not sure, but I think that we have this ability already, and the ram just isnt fast enough.

[AthlonXP1800]

Sound much the same thing with laptops with integrated graphics chipset that sharing memory borrowed from system RAM, like my laptop, it shared 64Mb of system RAM, the performance in games were really terrible slow. I think ATI wasted developed HyperMemory for nothing. :rolleyes:

[Ninja Prime]

Sound much the same thing with laptops with integrated graphics chipset that sharing memory borrowed from system RAM, like my laptop, it shared 64Mb of system RAM, the performance in games were really terrible slow. I think ATI wasted developed HyperMemory for nothing. :rolleyes:

I think you're a blind fanboi. Hey, that's just me though...

:rolleyes:

[retsam]

now that i think about it this sounds like the tech for xbox2 unified memory .....

[Cota]

UMA marketing jargon

[Sazar]

sounds interesting for lower-end products that don't require ultra uber on-board memory...

cost savings are always appreciated by OEM's... how practical and feasible all this is is what I wanna know :)

whens the first demo's ati :D

[nutball]

Sounds like virtual graphics memory to me, which 3DLabs have been doing for years.

[Dazz]

Sound much the same thing with laptops with integrated graphics chipset that sharing memory borrowed from system RAM, like my laptop, it shared 64Mb of system RAM, the performance in games were really terrible slow. I think ATI wasted developed HyperMemory for nothing. :rolleyes:
Yeah but the bandwith of AGP 4 & 8x is alot smaller and also system memory is getting quite fast now with most systems using 400MHz dual channel memory now. Also the least improtant data is being stored in system memory so it won't slow down as much.

[NightFire]

So, essentially, HyperMemory(tm?) is just an advanced and very effiecient way of sharing system RAM with the video card?
I don't know much, but couldn't that degrade preformance if you didn't have enough system ram? Or is ATi going to put a minimum required RAM for all of it's cards having that feature?

Sounds interesting...

[jbirney]

Yea I wonder if its aimed at the intergrated market?

[Remi]

That looks a lot like the virtual memory pushed by ms for the next DirectX. It seems ATI wants to make the first public move. Good for them! :)

I just hope they'll be more lucky with that first silicon implementation than with f-buffers.

[Soylent]

Alot of people seem to be taking this as some kind of thing for low-end/integrated cards.

Lets say a new model pops up in the horizon in some game. This is a very detailed game and the largest mipmap is 2048x2048 or 1024x1024 or whatever(not an insignificant chunk of GPU RAM). AFAIK currently games load the entire texture and it's mip maps all at once. Do we really need them yet?

Hell no, we need tiny mipmaps when it first pops into view, loading those huge ones will unnessecarilly eat extra GPU RAM and it will most likely drop your framerate/stutter when loading the textures depending on if it tries to precache before the thing comes into view or load the entire thing when it first becomes visible. What we really want to do is load the small mipmaps first and load the larger ones when we need them. For speed we want to minimize HDD activity so try to store this in RAM and transfer it to the GPU as needed.

We can't transfer it back from the GPU to the RAM over the AGP because it's very slow in that direction, so instead it's just overwrittern when we need more GPU RAM.(PCIe solves this)

Fixing these problems is important/beneficial for ANY graphics card. Direct x next is going to introduce virtual memory for graphics cards, which is similar but probably more advanced and far reaching than HyperMemory, AFAIK it's going to apply to shaders and stuff too and not just textures.

[jolle]

Sounds like virtual graphics memory to me, which 3DLabs have been doing for years.

Bingo, both Carmack and Sweeny has been pushing for that for years..
Makes even more sense on PCI-Express..

If its similar to 3dLabs VM then it will use onboard RAM as a cache memory and System RAM for storage, with a good protocol that handles it and makes sure things flows right and gets stashed at the right RAM it could be pretty good.
Not the "brute force" way AGP texturing seems to be trying to do, on a weak bus.

The reason Carmack and sweeny was trying to get ATI and Nvidia to get on that train was cause it would:
1. Free developers from the limitations set by Onboard RAM when it comes to textures and such.
2. Makes graphics cards cheaper, not needing tons of onboard RAM.

how efficient it is, remains to be seen to.