Hacker News Reader

Comment by ck2 | Hacker News Reader

How long until they just use dimm slots?

Hopefully never. Flash's characteristics are so different from DRAM that flash DIMMs just confuse people.

But there is http://www.sandisk.com/enterprise/ulltradimm-ssd/

You can use the DIMM slot for the DDR bus interface, but not used in the traditional way DRAM would be. The DDR interface is extremely fast, and can be purposed for uses other than the standard row/column/select/precharge etc commands.

You would likely need chipset support for it to be used in this fashion, though.

That's what flash DIMMs do. The problem is that if you stick something in a DIMM slot people expect it to behave like RAM, not like an I/O device.

I don't think the kind of person who buys a DIMM SSD is the type of person who is confused by that sort of thing.

To a non-technical person, mSATA, DIMM, and PCIE SSDs all look the same. These products aren't for them.

Agreed.

It would be nice to see a set of slots for both RAM or an SSD. So your laptop may have four slots total, and you can mix and match RAM or drives as you see fit.

The technology to do this is already commercially available in servers. Just need it to come to laptops or PCs.

At some point, I hope to see the lines between RAM and SSDs dissolve, such that the OS can just use a portion of the SSD for RAM.

Uhh, unless SSD technology changes significantly that line is a long long way off. SSDs are orders of magnitude slower than RAM. For example assuming a 2 GB/s speed is legitimate, RAM is over 22 GB/s. Plus RAM can be written and re-written without any wear, whereas SSDs have a max-write limit.

Right. Plus when you power off RAM ... Oops!

I think it's supposed to happen with these: http://en.wikipedia.org/wiki/Memristor

HP is working on a new architecture based on computers with just a buttload of persistent RAM, but so far it's vaporware.

Supposedly they will never sell this computer, only renting access as a "fast cloud".

Isn't that called "swap"?

No. Swap or paging is unused RAM "saved for later" (to free up real RAM space). It is never accessed or modified directly from processes except to push it back into RAM when a process requests a memory region which has been paged.

I would think the recent historical tendency is to have more cache levels, not fewer. This would only happen if the cost of storage loses its sensibility to speed.

SATA DIMM slots are already available:

http://www.vikingtechnology.com/satadimm-ssd

That only uses the DIMM slot for power - it still uses a SATA connection. This definitely has its uses, though.

http://www.sandisk.com/enterprise/ulltradimm-ssd/

I think this one meets his description and makes actual use of the DIMM for memory transfer, though I've done relatively little research on it.

Ugh. Well, I suppose it might be OK as long as it doesn't inherit anything about the DRAM hardware interface. Seriously, if you haven't seen how it works (test: are the timings tCAS, tRCD, tRP, tRAS familiar?) you are dramatically underestimating how ugly and restrictive the interface is. It's completely inappropriate for anything but DRAM and probably inappropriate for modern DRAM too. If not, it's probably because backwards compatibility with the interface has constricted DRAM design.

Speaking of which, I'd love to read a modern analysis of chip-to-chip serial vs parallel if anyone has one handy.

It may happen, but not without potential pitfalls. For example, modern CPU's cannot actually address the full 64-bit address range of memory, and if your drive is a DIMM device it is necessarily memory-mapped. CPU's can address a few tens or hundreds of TB. While few computers have that much disk, we aren't far off.

http://en.wikipedia.org/wiki/X86-64#Canonical_form_addresses

There's nothing in that article that says you can't write a driver which utilizes certain addresses as registers and other address ranges as visible sections of a mass storage device, for example. That's just one hack workaround I can think of that would allow you to put flash into a DIMM. You could combine that with DMA to read data out of your flash DIMM and into your RAM DIMM.

At that point, reading data would be a matter of setting up your paging registers and then queuing a DMA operation to copy data out of the exposed segment of your mass storage device and into RAM. Your flash DIMM would then be just another memory mapped hardware device.

Hmm, good point. Looks like I was overthinking things.

Reply on news.ycombinator.com