VRAMFS: Using Your Video RAM As A Linux File-System

[vmicho]

Ramdisk cards already exist... and would be far faster probably a bit quicker than an SSD.

Remember GPUs are not optimised for quick transfer to and from but for fast access locally between VRAM and GPU.

Well, in fact, they were the first (or one of the first) SSDs in fact. They were here long before the flash SSDs.

http://www.storagesearch.com/ssd-ram.html

Here just a short quote from another site:

Here's a simple rule of thumb based on analyzing published price data - for identical storage capacity - across a wide range of commercially available SSDs.
* RAM SSDs cost 9x more than SLC flash SSDs
* SLC flash SSDs cost 2x to 6x more than classic MLC flash SSDs (2 bits per cell MLC)
* classic MLC flash SSDs cost 1.5x more than 3 bits per cell MLC SSDs

[Tired_]

I wonder if this would have any potential for a home theatre pc. I have a video card in mine that has a lot of video ram (2GB, I think), but that is only used to decode video. If I used this to load the video real quick off the drive and into the video ram, would it make any difference in how the GPU decoder gets it? Does decoding video use a lot of video ram?

[wizard69]

Call me pessimistic,

Well that is kind! My first thought was "what in the hell are these?". Well you get the idea.

but this seems like a waste of time, although it's their personal project. Maybe if we start seeing PCIe based RAM expansion cards or a PCIe card that is nothing but boatloads of VRAM, then we could make good use of a VRAMdisk driver. As stated in the article, capacity is the problem. If I really need a RAM disk of any sort, it currently makes more sense for me to invest in a triple or quad channel mobo than anything else.

If it is a personal project I'm not really concerned at all. They are free to do whatever they want. On a side note expect to see GPU boards with 16GB of RAM soon. The capacity issue isn't a problem for every potential user. Why 16 GB, to compete with the coming Xeon Phi cards as compute resources.

I think the big problem here and what has created your pessimism is the costs of these RAM disks. There are simply better ways.

[Luke_Wolf]

Know of any off hand? I got spare RAM!

Also, if there's any USB based RAMdisk "drives" that can be attached to a Raspberry Pi or such, I'd love to know.

http://www.tomshardware.com/news/Super-Talent-Ram-Disk-RamDisk,23173.html

[schmidtbag]

The mid-2000s called they want their RAM prices back, that's 8GB territory and if you want 16GB of RAM it starts at ~$120
http://www.newegg.com/Product/Produc...size=30&Page=2

oh whoops I meant 8GB, my bad.

[STrRedWolf]

Ramdisk cards already exist... and would be far faster probably a bit quicker than an SSD.

Know of any off hand? I got spare RAM!

Also, if there's any USB based RAMdisk "drives" that can be attached to a Raspberry Pi or such, I'd love to know.

[Luke_Wolf]

4GB of RAM is roughly $70-$100, depending on how good it is. If you were on a tight budget, the money you could have spent on that RAM could have been spent on a better GPU, which in turn, may include more than 2GB of VRAM.

The mid-2000s called they want their RAM prices back, that's 8GB territory and if you want 16GB of RAM it starts at ~$120

http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&N=100007611+50008476+600006050+4027+4017&IsNodeId=1&name=%24100+-+%24200&Order=PRICE&Pagesize=30&Page=2

[schmidtbag]

I have heard about this project many years ago, but its interesting to see they want to take advantage of a simulated RAID0.

I think what would be more interesting to me is to see how much of the GPU core is required for this. The reason I'm wondering this is because you could build a PC with maybe 128MB of regular SDRAM and have maybe 4 2GB or 4GB GPUs in PCIe slots. These GPUs could be used for something like folding or SETI, which really take up almost no RAM at all. What you really need are the GPU cores. So, you could instead grab maybe 1GB from each GPU and use it for the rest of your system.

4GB of RAM is roughly $70-$100, depending on how good it is. If you were on a tight budget, the money you could have spent on that RAM could have been spent on a better GPU, which in turn, may include more than 2GB of VRAM.

[ales-76]

What's the usecase scenario?

I understand this might be interesting experimental feature, but I can't see a usecase for it. Neither for swap nor for VRAM backed file system. I mean RAM is cheap, if you are running out of available memory you can buy additional RAM for measly dollar so need for fast swap is virtually nil. On top of that one can simply create compressed RAM backed swap (like zswap or zram) or RAM backed file system (like RAMFS or TMPFS) if required. That would be easier, faster and cheaper than buying expensive GFX card with tons of VRAM and then use it as storage device. Plus I guess one would run into PCIE bandwidth limit when runnning GFX demanding stuff and swapping into VRAM simultaneously, slowing down both, thus negating any benefit.

[nigeil]

While this does sound like a solution looking for a problem, might this find more traction in render farms? Although I bet 50 cents their render time is pretty full on, and at a price premium. Maybe in science labs as well?

When I first read about this project my mind jumped to some of the computational science problems that might benefit from it. Scientists on the whole are still just starting to take advantage of GPGPUs, and regardless of whether or not this filesystem is actually used, it's nice to know that it's there! I say rock on; doesn't hurt anybody.