Ddram disk – A PCIe card with 14x RAM slots

Something seems a bit dodgy. The about page shows the top and bottom PCB layouts without any RAM sockets, only dozens of DDR RAM ICs soldered directly to the back of the card.

...Which would be fine, had I not stumbled upon (what seems to be) the original project[1], in Russian no less.

The blog shares the same block diagram, images and DDR PCB layout. There are no DDR RAM slots. Further, the slots are through-hole and not SMD, you wouldn't be able to attach a heatsink directly to the back of the card as shown in the images.

[1] https://habr.com/ru/post/567742/

>using refan DDR memory...Kllisre, ZIFEI, TANBASSH, Rasalas etc. use the refan chips

"refan" I think they might be talking about factory reject recycled/repaired ram. Chuck Peddle (6502) had some patents for this and ran operations in Asia for a company doing this on a large scale "Memory module assembly using partially defective chips" https://patents.justia.com/patent/6119049 https://patents.justia.com/patent/RE39016

Oral History of Charles Ingerham “Chuck” Peddle https://archive.computerhistory.org/resources/access/text/20...

>Peddle: I had a patent on how to use partial DRAM. And we turned that patent into a very large business making DRAM in Sri Lanka, in India, using reject die from Micron.

>Peddle: We buy the good ones in the front, probably about half price because they sell them about half price. Although we buy them less than half price because a lot of them don't make the 50% numbers that other people are looking for. And then the ones in the back, we basically almost get free because we can use them up.

That business died when Micron stopped selling partials(reject dies) to anyone but their subsidiary SpecTech. "SpecTek began at Micron in 1988 as a component-recovery group. In the two decades since its inception, the company has grown from an internal group to a manufacturing division; that experience has produced a product portfolio and reputation that makes SpecTek a leader in application-specific memory solutions.". Every time you buy Crucial you might be buying inferior SpecTek factory remanufactured/patched from defects product.

> We lost 4,500 jobs in a weekend when they stop selling partials.

Nowadays a TON of Chinese brands are into partials. Cheap flash memory was first - why throw away 90% defective flash die when you can make small SD card/USB drive out of it. I have no doubt they moved to ram now.

For the rest, in current chip shortage $280 wouldnt even cover 4 FPGAs pictured on that website :)

I'm also doubtful that a memory controller with the necessary pin count can even fit under that heatsink. 14 DIMM slots can't all be sharing just one or two 64-bit channels at anything approaching standard speeds and timings.

Apple solved this problem at scale some time ago. Im not sure slamming ram on a pcie card is a better solution than including a piece of effaceable storage somewhere.

Write endurance I guess would be one.

Samsung 980 Pro is under warranty up to 600 TBW.

I don't know if that is external data written, or also includes whatever overhead the drive's internal processes have, which I imagine is higher when you run it near max capacity.

RAM has virtually unlimited write endurance.

I was interested for CCTV recording, as it can be murder on standard HDDs and consumer SSDs. Continuous writing at whatever the total bitrate of all cameras is, plus whatever index you're using, plus whatever clips its generating.

Ideally HDD for the constant, SSD for the clips and index.

I had a small (120GB) clip and index SSD drive start to struggle in a dirt-cheap non-critical system and found it had written 42TB in a year. Now part of that was poor configuration, but the drive was REALLY starting to struggle, throughput would occasionally drop off a cliff.

The system was battery backed, so a RAM disk would have been fine, and on triggering the UPS I'd have it just copy the contents to a fast HDD.

Speaking as someone who actually runs NVMe RAIDs for large-ish databases, the NVMe speeds quoted are usually only attainable for large sequential streaming loads. While they are not spinning disks with heads that need to seek, there is still very substantial random I/O overhead with SSDs, including NVMe. Even with a top of the line model like a Samsung 980 Pro, you might only get 5% the PCIe 4 bus speed with heavy random IO performance.

RAM has cachelines and such, but postgres reading 8k at a time isn't limited by that.

How to work something like this into Linux as say, a very fast swap device, is another question.

RAM doesn't have a fixed number of write cycles. It can be written to indefinitely, but NVME flash has an estimated number of TBs that it can write before it fails.

RAM is volatile (the data is destroyed when the power is cut), which can be a security feature. There appears to be a battery backup on these boards, but I imagine that it can be removed easily.

Aside from what everyone else has stated, a ram disk like this seems to have a very very good rand4k speed, so it would be extremely useful for buffering high volume small size, erratic data loads while decreasing the number of writes to long term storage.

Latency maybe?

Looks like the card is PCIe 4.0 x4, which gives about 8GB/s. Not sure if this board is driving the RAM at full speed, or how many controllers there are, but more lanes could make this faster.

I haven't kept up with the speed and feeds, but in general PCI bandwidth is a decimal order of magnitude less than all the memory controller bandwidth

And that is the killer app! When you have 3 dozen 4GB DDR3 SIMMs sitting around doing booger all, it is overwhelmingly tempting to get a card like this, if it is priced aggressively, and run something really fast. Re-image the OS from slower storage at boot, or whenever the 18650 falls out.