I'm not seeing much of a benifit of using a DIMM/DDR interface over SATA III, PCIe 3.0 16x, or similar.
The most obvious is much lower latency. It's very close to the CPU and uses the same controller that RAM uses. I'm also suspecting that they are also off-loading to CPU and the RAM controller all the wear-leveling stuff, and all things that are done by a SSD controller in general.
A modern CPU won't really notice this tiny load but it is again another latency-cutting trick.
For a server that has hundreds of requests to fullfill per second, low latency does make a lot of difference. For a consumer PC it still does if you have too high latencies, but this would be way beyond overkill.
One of their two possible uses (as from the press release, and they have made drivers that allow this) is using them as cache. The second is plain
FAST SSD storage.
The flash cache is a trick that Android devices do all the time. When something is idle in RAM for a too long time (parameters for this vary) but is still likely to be useful in the future, instead of flushing it out of RAM to free up space for the task at hand it gets written in the cache partition (on the flash chips).
Since Android devices are tiny and all the stuff is run by their SoC (a single chip that contains CPU, GPU, controllers for screen, storage, USB, and whatever else), they get some pretty good latency and thus are more performing than their specs would suggest.
The same is done by hybrid drives. But in this case it's stuff taken directly from RAM, not data accessed from an HDD (that would need time to unfold in RAM).
Unless maybe flash memory is capable of much higher speeds than I'm aware of?
That depends from implementation. You know RAID 1? They could have done something like that.
Since this is a server-oriented product, reading speeds are far more critical than writing speeds.
Still I frankly don't know the actual speeds of flash chips alone at the moment. I suspect that the controller's own speed makes a difference.
(The GDDR5 throughput on my SLIed PCIe 3.0 16x graphics cards is roughly 143GB/s)
These things are going to be DDR3 or at most DDR4. Max transfer rate for a single DDR3 bank (at 2133 Mhz) is around 17 GB/s, don't know about DDR4 but I doubt it will go beyond 50.
I doubt flash chips even in a realistic DIMM-sized RAID 1-ish implementation can get beyond the 3 GB/s, but more bandwith will be certainly useful in case there are multiple file transfers done at once. (again a server does experience this pretty often, consumer PC do too, albeit at a smaller scale)
And then you still have the limited read/write cycles compared to RAM.
This is relatively irrelevant. Modern flash chips can be written on so much times that even servers can use them.
The price tag of a single server-grade SSD is usually well above the price of a fully pimped consumer gaming rig, though. Then again server farms are murderously expensive anyway.
Seems like more of a niche solution that most won't get much benefit from.
It's server hardware. And as said above in post 5, a lot of servers don't have the physical space to fit a PCI-e SSD nor the space to put a massive RAID array. But all servers have a ton of RAM sockets, that in the case of 1U or blade servers are rarely fully filled with RAM.
lehnerus2000 said:
It could be useful in disposable devices like phones and tablets.
Afaik, all mobile devices solder the RAM and storage chips directly to the "main" (and only) board.
being this the same flash chips used elsewhere I don't see the point. The ReRAM chips will do a lot of difference.
Btw, If someone combines ReRAM and this technology and releases it at a decent price it will blow away all SSDs on the market (PCI-e or not). As those chips promise to be faster than flash too, and will take full advantage of the gigantic data transfer speed of the DDR interface. I mean a single slot has the same oomph of a full x16 PCI-e 3.0 slot, and when DDR4 comes it will beat the 4.0 too.
