[1] conventional single motherboard designs that is, larger of the top500 machines seems to be at the single petabytes now

[2] https://news.ycombinator.com/item?id=9582023

Some back of the envelope math -- or math in your head if you are good at that sort of thing -- put the price of 1.5TB of RAM at a cool $18,000.

If my memory serves me well, I routinely bought 1Mb RAM modules for 5.500 Spanish pesetas (33 Euro). This was around 1987-1990 I believe. If my math is right, that would be more than 4 million Euro for 128GB RAM.

1994 prices are US$30.0/MB or $8K for 256MB of PC-grade memory. Looks like the DEC machine used 200-pin SIMMs, says https://www.memoryx.com/ms15da.html (listing 64MB for $200). Don't know how that compares.

BTW, $400 for a DEC Alpha 3000/500 at https://www.ebay.com/itm/DEC-DIGITAL-ALPHA-SERVER-3000-500-P... .

To be fair, you have to factor in Moore's law. I'm not great at math, but I think you'd still end up with something like half a million. $18k for 1.5TB is still a steal.

This is not an accurate calculation, first, the main cost of an Arduino is its PCB and the supporting electronics, plus some profit margins to cover the overhead for the manufacturer.

Also, the RAM is not "installed", the actual device where the 2 KiB RAM resides in, is etched on the same silicon chip of the Atmega328p microcontroller. A 328p only costs $2 each, and close to $1.5 in huge quantity [0]. Since it's pretty much a standalone device, needs nothing (not even an external crystal) but power, some wirings and two decoupling capacitors to operate in a minimum system (and yes, even this configuration can run standard Arduino code). Let use a conservative number, $3, it would "only" cost you about 2.4 billion dollars.

Furthermore, Atmega328p is marketed at a "mid-end" 8-bit microcontroller offered by Atmel, mainly for its GPIOs. If you use an ATTiny1614, which is basically the same hardware with much fewer GPIO pins, it would only cost $0.7 for each, let's use $1 as a conservative number, then it would only cost you 0.8 billion dollars for 1.5 TiB of RAM, which should already be considered a very impressive number, since you couldn't even get a NAND gate, or even a MOSFET in the late 1970s for less than one buck. If adjusted for inflation, today's price would be as low as ~$0.17 per 1024 bytes.

And then you consider the fact that it's not only a RAM, but in fact a computer... Isn't it a miracle of Moore's Law?

[0] https://www.mouser.com/ProductDetail/Microchip-Technology-At...

It is true that those micro controllers are bundled with the memory. In other words, the memory also has computing power installed. A capability that some modern memory chips also include. You may argue that we should compare with those memory chips, however, in the case of micro controllers we cannot separate the memory from the computing unit and we need the computing unit for communication. Is the comparison thus unfair? Maybe.

I should have compared with magnet core memory used in the Apollo program as an example. My gut tells me that we don't have enough rocket fuel to install that kind of memory in space.

AMD's new epyc supports 4TB without artificial price increase, IMO some healthy competition for the server market.

I'd like to see how the flash-as-RAM devices work out for some of these apps, i.e. is it purely single-threaded random access or can it be batched/cached?

But "gamers" who like to overclock, and AMD users who use motherboards that don't generally support ECC don't care.

For professional work, I think you'd be crazy not to. You get 1 bit-flip per year per gigabyte. So for my 128GB system we use on desktop configurations and build servers we'd get 2 bit flips/week. Do you really want to send a build out to a customer that has a potential bit-flip in it?

That's only about three times more expensive than if you'd extrapolate from regularly sized DIMM modules.

Also, 1.5 TB RAM became somewhat commonplace around 2014-ish. LinkedIn had to have Dell custom build them a 1 TB RAM machine around 2009. So 10 years later for the same build to be passé is fully reasonable when a lot of the market is doing massive or very latency sensitive graph calculations that won’t accept the variances of distributed algorithms.

Being able to reach 12 TB+ with something like Optane is much more likely than with RDIMMs though, and memory latencies would probably get insane enough that Optane would start looking viable anyway.

Well, less than any other time in history. Not sure I get the point here.

Unless you truly need DDR4 speeds, I would think you would be better off for your money using CPU Raid and NVME. More drives, more speed. One guy used 8 drives and got 28Gbps.

https://i.imgur.com/AuCWq3r.jpg

second hand 64GB DDR3 ECC REG RAM is about $100 each, a dual socket c602 motherboard such as my Z9PE-D16 has 16 memory slots, having 16 x 64GB = 1TB RAM in such a workstation costs you $1,600.

bargain!

On my personal desktop right now. I have 20gb of ram and over 250tabs open on Firefox, spotify, VCode, Postgresql, pgadmin, multiple evice & fbreaders, rabbitMQ, some nodejs servers, 2 linux containers, 1 docker container, mysql server, 10+ xterm sessions, a few screen sessions and I'm using 14Gb.

Then you can get rid of all disk reads for normal operations. I imagine gamers would absolutely love it, particularly if creating and populating a ramdisk were made really simple.

Its unclear what is this article about, yes large modules are expensive, but in DDR3 era 128gb dimms did not exist. You're paying the early adopter tax. If you want to get huge-ram systems for cheap, there are a lot of used 4-cpu servers, the whole system can be assembled for under $5k.

It’s like Apples new Mac Pro and display. Those prices are just dialed that way to keep them out of home offices. It’s infuriating, yes, but prevalent.

To my disappointment, it goes completely unused more often than not. According to Activity Monitor, I currently have around 7 gb just sitting idle.

Typically the only time I use enough memory that my working files don't fit in the page cache along with the application memory, is when I build Android.

It's possible you just don't have that much on your disk that you use.

At all times only a very small fraction of my ram is free.