Reply to post:

Everspin's new gig: a gig or two of non-volatile RAM on PCIe


> The first of the new “nvNITRO E” range will be a half-height, half-length PCIe card that can operate as an NVMe solid state disk, or as memory mapped IO (MMIO).

How is this different to other NVMe setups? I have a PCIe NVM card (120GB) in my server. It cost me £70 all in, and is rated at 6Gb/s bulk transfer and some stupidly high IOPS that I can't remember right now.

I can use it as swap (in which case it just becomes allocatable memory, and the OS handles all the paged MMAPing), or I can use it as a file store, and MMAP files directly on it for the same effect.

I don't see what is special about this startup, except their NVMe offerings are really low capacity? The 120GB SSD is running as swap and it actually works pretty well. I have used ~105GB of swap on the 32GB RAM machine, and it was still usable and churned out data at an acceptable rate (this was for peak loading, most of the time 32GB is enough, I just didn't want the machine to die when the peaks come in, and couldn't justify the cost of 256GB of RAM for it ).

> Everspin asserts its product is rather faster than Intel's 10 µs and, critically, that you can read and write to it all you like without the prospect of the medium degrading.

That will depend on the cost. When my SSD eventually wears out, I will just buy another one (as they would most likely have gotten cheaper by then) and carry on.

It might be cheaper to just treat the SSD as consumable, and replace when they wear out. Each time you replace you will get a newer/faster/cheaper/higher capacity version due to the march of technology.

> The cards claim 1,500,000 IOPS with six microsecond end-to-end latency, making them rather useful in scenarios like high-frequency trading where the odd microsecond can be the difference between profitability and purgatory.

HFT shops have long since moved away into FPGAs with local RAM, computers have been relegated to babysitting the FPGAs and monitoring/restarting them as and when needed. You don't need uber low latency memory for that.

They do mention merging their memory with FPGAs, and that might prove an useful niche for the technology, but that hasn't been done yet, and no idea if it is a worthwhile and profitable niche (FPGAs do need to store some data, but not much, most of HFT is simple arb strategies just done stupidly fast based on data in/out of the network port)

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2022