Intel and memory standards
I still haven't forgiven them for that Rambus RIMM nonsense. Not sure I ever will.
Gold plate can give a durable and affordable alloy a 24-carat veneer finish, adding value to cheap metal. DRAM gives Netlist's Hybrid DIMMs a cache veneer, providing what looks like DRAM to applications but is really persistent NAND underneath, cheaper than DRAM and lots of it. HybriDIMM is the result of combining Netlist's …
"Say hello to Netlist's flash-DRAM grenade" is right, Chris! What a way to pummel Netlist's only competition, Diablo Technologies Memory1. Memory1 can never compete with the Samsung/Netlist HybriDIMM. The numbers are nowhere in Diablo's favor. Give Diablo an "A" for their cunning effort and thievery, but given them an "F" for execution. WTG Netlist!
Your enthusiasm may be a bit premature. We're talking about whiteboard specifications, not production measurements. All the numbers are PR-driven and subject to verification with actual, real-world results.
We've been promised the moon before, many, many more times than we've ever gotten it.
I doubt you can have performance AND low cost AND no endurance problems.
A RAM chip can't say "hang on a minute, I need to sort out my fragmentation problem". It must deliver the data at however many clock cycles it is after the command.
So I think this appears as a (memory-mapped) storage device in a DIMM slot rather than an M.2 or PCIe slot and cannot be used as RAM. Obviously unlike normal storage devices can't do interrupts either so lots of polling needed.
Clearly you have no clue what HybriDIMM actually is or you are just an unwise skeptic. HybriDIMM operates at nano DRAM speeds. Not NAND speed. Not all data is needed at once hence the software developers embracing Netlist's SCM. The CPU sees the first step in this architecture as RAM. Not "memory-mapped storage".
Netlist developed the HybriDIMM together with Samsung and demoed the product to Samsung engineers. Do you think Samsung would align with Netlist if performance (aka endurance) and cost were an issue.
By de-fragmentation I assume he was referring the the host of issues around bad block management and write amplification that occurs in SSDs. Those problems are very real and a large part of getting an enterprise grade SSD working. If this isn't done right, this product will not survive write intensive workloads and it won't last long.
I saw the Netlist Demo at FMS, and it was weak. No real application running because the presenter admitted the drivers aren't ready yet but will be later. I guess there was a push to show "something" at the show and at least its not complete slide-ware so that is great to see.
So ignore the hype for now and lets wait and see how well the driver works. Then we can see what real applications can do with this hardware. Then it will be interesting to compare this solution with 3DX, Memory1 and other NVDIMM offerings on head to head benchmarks. That would make the next FMS really interesting.
What isn't clear to me is whether for this to be effective all DIMMs in the system have to be HybriDIMMs or whether mix and match works. In other words, will the driver stack be able to move data from the flash to the local DRAM and then move it to where the application needs it. A lot more information on the software stack and how it fits into real applications would be very interesting.
Biting the hand that feeds IT © 1998–2020