What kind of RAM do supercomputers use?

I was reading an article about one of the worlds fastest computers and this is one of the lines which interested me:

To mimic this relatively minuscule amount of brainpower, researchers used the Fujitsu K to connect a total of 1.73 billion virtual nerve cells by 10.4 trillion virtual synapses (with 24 bytes of memory in each synapse). In total, this added up to around one petabyte of memory, which is the equivalent of about 250,000 standard PCs. And remember, all that's still just one percent of what your brain does every single day, in the time it takes to blink a few times.

So this implies that those supercomputers use normal desktop memory but with much more quantity (250k times more) than a normal desktop?

I read this popular question here on SuperUser. It tells us also something about SRAM - the RAM that CPU's use for cache because it's much faster than normal DRAM.

So I looked here and this line seems to be interesting:

The only reason memory sticks don't use SRAM is the cost. Instead of paying $50 for that 256MB chip, you would be forking over $500. Ouch.

So SRAM modules are possible (and I won't hide the fact I'm really interested in buying 8GB of SRAM memory for my laptop... but can't find anything)

But now the real question: Do supercomputers use SRAM for their memory or just normal DRAM? And if normal DRAM, why don't they just create SRAM memory for supercomputers, supercomputers cost hell of a lot so what difference would that few extra million bucks make?

Given that many supercomputers today use server-grade x86-64 processors, like AMD Opterons, they would probably be fitted with ECC memory, most likely DDR2 or DDR3 SDRAM.

ECC memory has the ability to detect and correct most memory errors. It is generally not seen in consumer computers, but is commonly found in servers and workstations, where maximum reliability is required. An example of this kind of memory can be found here.

ECC support in consumer computers is rare, but note that AMD FX processors fully support this type of memory, and with a proper motherboard, the enhanced reliability of ECC memory can be realized. ECC memory can be expensive, though, and tends to be marginally slower (by 1-3%) than non-ECC memory.

Some newer supercomputers use GPUs for computation alongside conventional processors (as in the Cray XK7). These GPUs generally use GDDR5 memory.

The design of the Super Computer decides the type of RAM used.

For example, a Beowulf Cluster will likely not have some exotic form of memory because:

A Beowulf cluster is a computer cluster of what are normally identical, commodity-grade computers networked into a small local area network with libraries and programs installed which allow processing to be shared among them.

The same goes for many of the worlds super computers, where they are built essentially with off the shelf parts, Intel processors, maybe NVidia GPUs for parrellel processing.

Feel free to peruse the worlds fastest computers, and the common trend is to use common parts in uncommon ways.

It is likely due to cost, because at the scale of these computers (20,000 nodes perhaps) that difference of $450 per 256MB you quoted adds up ($9,000,000 if each node only had 256MB). This math seems outdated to me, but the point remains it matters.

Most Super Computers that use x86 or x64 processors (Generally Xeons or AMD Opterons) will use ECC DRAM. It is a little slower than desktop memory, but the data is more reliable.

If you count the RAM on the GPUs that are used, then GDDR5 is another common RAM in Super Computers.

But, SRAM and any other exotic RAM is not common in modern Super Computers.