graphics RAM

Hi /prog/

would it theoritcally be possible for an OS to make use of the RAM on a graphics card as system RAM, eg at boot?

If the said graphics card supports mapping its own RAM at a given physical address, then I guess it's possible to map those pages at a given virtual address and then use them as if they were RAM, even running programs from them.

I see, I mentioned the concept to someone and he was a total douche about it ('gpu ram design is optimised for vector parallism, your concept is dumb because caching structure, controller, bus, timing, GART tables, etc')

in fairness i don't know about these things but i still thought it had potential...

The reason it came to my mind in the first place was because i happened to be reading about the next radeon lineup and noticed that they say:
'Support for x86 addressing with unified address space for CPU and GPU, 64-bit addressing'.

Now i don't actually fully grasp what that means, but it seemed to me that gfx ram being faster than system ram it might be a good way to speed up boot times/improve performance of 2D apps/when gfx RAM isnt full.

Can i take it to mean that the next radeons will indeed be able to map their RAM as you describe?

http://en.gentoo-wiki.com/wiki/Using_Graphics_Card_Memory_as_Swap
https://wiki.archlinux.org/index.php/Swap_on_video_ram

Now i don't actually fully grasp what that means, but it seemed to me that gfx ram being faster than system ram it might be a good way to speed up boot times/improve performance of 2D apps/when gfx RAM isnt full.
It is possible, but it will never be as fast as the central RAM when accessed by the CPU. It might serve as a better swap space than the HDD, though.

'Support for x86 addressing with unified address space for CPU and GPU, 64-bit addressing'.
I believe they're going to implement what nVidia call bindless graphics, e.g. letting the client application access some parts of the VRAM bypassing an important part of the driver. However, that's only useful for running things on the GPU.

>>1
The problem is generally that it takes less code to initialize the DRAM controller on a motherboard (or inside the CPU, as it often is the case in newer computers) than it takes to (1) initialize PCI (2) find the graphics card (3) tell the graphics card to map its RAM at a given physaddr.

I don't know if this trick is still used but back in the days of cacheless CPUs the BIOS would detect the video adapter and use part of its memory for scratch RAM because it wasn't sure how much system memory was installed yet.

>>4
>Gentoo wiki
>Arch wiki
You just have to love these guys.

cool thanks for the replies, i'm glad to see that this is already possible.

>>6
i take it the CPUs with GPUs on-die use system ram for graphics right? i guess if intel/amd designed systems so that you could add vram in the same manner as system ram it would kill the gfx card industry

>>7
do you have any links with info/history of this technique?


do you guys have any opinions on if/when SSDs could make system RAM redundant?

my opinion on the question of when SSD will replace RAM is that it's the dumbest thing ive heard in a while.

>>8
We are free software hackers, and this is simply the natural progression of communities with software freedom.

Of course. Linux doesn't need video ram. It can run in textmode.
disable swap for performance and use that videoram for disk cache or ramdisk.
Also check out if your sound card has any RAM and use it too, since Linyx already has a PC speaker.




orbis terrarum delenda est

>>12
You can run Windows Server headless these days with just a text terminal or remote shell.

>>10
thats a very pretty comment, but maybe you could elaborate a bit?

>>10,14
SSD degrades in several years. It not meant for random access.
Try swapping on SSD or any hard work and it will start shitting out bad blocks.
I can't imagine SSD handling all the RAM streaming at DMA speed. Its way slower, block-oriented, and far more failure prone.




orbis terrarum delenda est

>>15
6/10. Nice usage of lies.

>>14
SSD is extremely slow compared to RAM.  PCI is very slow compared to RAM.  Why you want to use these things instead of RAM is baffling.  You want you're computer to run like a snail?

>>17
Enjoy your data-loss during a power failure without a UPS.

>>17
Enjoy your data-loss during a power failure without a UPS.

>>15
i wouldn't have imagined that there would be hybrid SSD/mechanical drives either a few years ago, but that hasnt stood in the way of progress... i'm not necessarily talking about hardware/techniques that exist today, rather the possible directions these things could evolve in/convergence.

>>17
i'm (i assumed obviously) not interested in using these things to slow the system down, rather to reduce the number of components in the system and also to make better use of them by combining their strengths.


essentially i'm wondering about the modulatiry computers enjoy, and how it will change in the mid-term future. My father often told me stories of how he would horse-shoe different engines into cars, much in the same way as we are able to change a CPU, and how technology has changed so that its now much harder to do, albeit with certain other benefits to counterbalance.

to my mind it makes sense that some components be integrated, and seeing as we are now in the many-core age it seems sensible that memory would be on-core, ie when you want to upgrade you add one package which contains a processor, a memory controller and RAM all in one. surely then you can kiss your gaping memory-hole goodbye? The alternative would be RAM [possibly in slots] on SSD seeing as i gather that getting data from drive to RAM is also quite slow. either way, it seems that RAM being isolated is archaic.

>>20
The memory and CPU may both be based on silicon, but the manufacturing process is very different. You'd also reduce your yields and make your chips a lot more expensive. You could get away with putting the memory in the same package. Some newer SoCs almost do this by using the package-on-package technique for their RAM, but they're also typically limited to 512 MB.

check 'em

>>21
it wouldnt necessarily have to be in the same chip, the design of CPUs with GPU/RAM could become similar to the design of gfx cards as they are now. it would probably make sense to have it slot based too so that you can easily add more.
when you want to upgrade, you just add one component, instead of a second graphics card and/or more ram, plus you get the added benefit of more processing.

>>23
You don't want to put your memory on the same bus as your peripherals.

>>24
i dont see why it would; i gather it isnt now, so why couldn't you just bring it on-board? the only things that aren't practical to have on the CPU are storage and expansion cards.

>>22
nice, dubbs!!!