Idea: Cheap ssd?

[Googer]

My thoughts are still in beta stages and definitely not a complete or deliverable product. However, I'd be interested in hearing your comments, ideas, or opinions.

I've been thinking about the ridiculous cost of SSDs, especially those based on DRAM/SRAM. And felt that manufacturers overcharge compared to other products using similar technologies.

For example:
Intel can sell a chipset with memory controller and SATA controller for about $30 including profit. Similar tech in ssd form usually costs well over $350 (excuding ram). Most of these devices use the SATA controller to feed their data to the CPU/RAM (EG Acard, HyperOS Systems, Gigabyte IRAM*) thus limiting their performance potential.



PCI-e PCI-L (legacy) based RAM SSDs use similar tech found on all motherboards and have been known to cost hundreds or thousands of dollars for a limited amount of storage.

If NVIDIA/ATI can sell graphics cards in PCI-e form with 2GB of HIGH SPEED RAM with Dedicated processor on a bootable add-on card for around $100. Why can't these "SSD Companies" do the same?

Well my idea is that someone (probably not going to be me) write a custom software package and/or bios for NVIDIA or ATi and use the GPU and RAM is a storage device for quicker loading of frequently used APPS like Firefox or a favorite game. Or offload the swap file on to the 2nd GPU-RAM.

I wouldn't advise anyone do this to their primary GPU but instead to purchase a low cost high ram capacity GPU to augment their primary Graphics card. Currently newegg.com has a 2gb graphics card for $84 and would make for a great test candidate. 2GB of high-bandwith storage with low cpu usage for under $100 sounds like a winning proposition to me.







*iRam costs less than $350 but has low capacity

 

[Blain]

Sounds like your thoughts are still in a beta stage. 🙂

 

[Googer]

Sounds like your thoughts are still in a beta stage. 🙂

What would need improvement, mass production?

 

[jchu14]

Wouldn't it be easier just buy more RAM and create a RAMdisk for that purpose? I don't know if a GPU-RAMdisk will be faster or slower, but I would imagine the performance difference would be hard to notice. A GPU-RAMdisk will be volatile so it cannot be used the same way as a hdd or sdd.

 

[Googer]

Wouldn't it be easier just buy more RAM and create a RAMdisk for that purpose? I don't know if a GPU-RAMdisk will be faster or slower, but I would imagine the performance difference would be hard to notice. A GPU-RAMdisk will be volatile so it cannot be used the same way as a hdd or sdd.

Sure but it would be more expensive.

Since most RAM-Disks use SATA and not PCI-e x16 as their interface you wont get as much bandwith or latency performance and you'll have higher CPU usage vs a PCI-e solution with it's own dedicated processor. A GPU-RAM Disk would cost much less and considering the bandwidth that GDDR 3, 4, or 5 can provide it is much faster this seems to be a better price to performance option.

 

[jchu14]

Sorry, when I say RAMdisk, I mean a virtual disk made from system ram, not dedicated sata/pci-expess ramdisk cards. There are software like dararam that allows you to allocate part of your system memory to act as a RAMdisk. I've used it before to make a RAMdisk to hold my brower temp file.

 

[Googer]

There are software like dararam that allows you to allocate part of your system memory to act as a RAMdisk. I've used it before to make a RAMdisk to hold my brower temp file.

I've used a few of those before. But they steal memory away from applications and rob the CPU of cycles (higher CPU usage), ram availability, and memory bandwidth. Plus there is also the cost of the RAM-disk software it's self plus any additional ram you'll need to install in your system.

 

[WildW]

Doesn't Windows do exactly this now? Superfetch in Vista and 7 does store frequently accessed programs in memory so they're opened faster. It even seems to work properly now in Windows 7.

 

[Throckmorton]

I've used a few of those before. But they steal memory away from applications and rob the CPU of cycles (higher CPU usage), ram availability, and memory bandwidth. Plus there is also the cost of the RAM-disk software it's self plus any additional ram you'll need to install in your system.

Yes, they steal memory that is cheaper than GPU memory. You can spend your money on a video card just to cannibalize its RAM, or you can spend your money on more system RAM for the ramdrive

 

[Googer]

Yes, they steal memory that is cheaper than GPU memory. You can spend your money on a video card just to cannibalize its RAM, or you can spend your money on more system RAM for the ramdrive

The Ram Drive software uses CPU cycles. Robs your apps of memory bandwith and increases CPU usage.

Using a Video card allows you to offload that process to your GPU and allows you to gain memory bandwidth.

Think of F@H, you have the option of a CPU or a GPU client. I am proposing a GPU client for a RAM-Disk program. Better performance.

 

[Googer]

Instead of buying a new one, this could be a good use for an older graphics card.

 

[superunknown98]

Sure but it would be more expensive.

Since most RAM-Disks use SATA and not PCI-e x16 as their interface you wont get as much bandwith or latency performance and you'll have higher CPU usage vs a PCI-e solution with it's own dedicated processor. A GPU-RAM Disk would cost much less and considering the bandwidth that GDDR 3, 4, or 5 can provide it is much faster this seems to be a better price to performance option.

This idea does not seem like a good one. The reason GDDR 3,4,5 has such massive bandwitdh is because video data is processed entirely(almost) on card. Once data needs to be pulled from system RAM you are limited to PCI-E X16 lane 8GB/s. That is slower than Dual channel DDR2 533.

 

[Zap]

The Ram Drive software uses CPU cycles. Robs your apps of memory bandwith and increases CPU usage.

I don't see that as much of a problem. I already have way more CPU cycles, memory capacity and memory bandwidth than I actually use/need (socket 1366) so no harm, no foul in allocating some of that untapped potential for a RAM drive. Heck, I was doing that with one of my socket 775 setups with dual channel DDR2 and a dual core, and I never noticed any performance hit (actually, didn't notice performance gain either 🙄 ).

I am proposing a GPU client for a RAM-Disk program. Better performance.

Maybe someone can whip up something in CUDA?

 

[Googer]

We'll it's not going to fit everyone and certainly is best used on your last generation card that you just replaced.

 

[Googer]

I don't see that as much of a problem. I already have way more CPU cycles, memory capacity and memory bandwidth than I actually use/need (socket 1366) so no harm, no foul in allocating some of that untapped potential for a RAM drive. Heck, I was doing that with one of my socket 775 setups with dual channel DDR2 and a dual core, and I never noticed any performance hit (actually, didn't notice performance gain either 🙄 ).



Maybe someone can whip up something in CUDA?

CUDA is very similar to C, so a lot of programmers should already be familiar with it.

OpenCL is similar to CUDA, however OpenCL isn't restricted to nVIDIA only. OpenCL runs on both ATi and NVIDIA.

I'd say OpenCL is the way to go.

 

[Googer]

This idea does not seem like a good one. The reason GDDR 3,4,5 has such massive bandwitdh is because video data is processed entirely(almost) on card. Once data needs to be pulled from system RAM you are limited to PCI-E X16 lane 8GB/s. That is slower than Dual channel DDR2 533.

8 or even 4GB/s sounds good enough for me and lets not forget that PCI-e 2.0 doubles that bandwidth.