Take the 2-minute tour ×
Super User is a question and answer site for computer enthusiasts and power users. It's 100% free, no registration required.

I was reading about memory hierarchy and I read that you can built a computer with just cache but it wouldn't be cost effective. I was wonderig if super computers are made with only cache technology?

share|improve this question
top500.org –  dmckee Oct 5 '11 at 20:30
add comment

3 Answers

up vote 1 down vote accepted

Perhaps its a misunderstanding that supercomputers are primarily about speed, because that's not necessarily the case.

Supercomputers are about computing, about power. Dealing with very large sets of data containing very large or very complex values very quickly. This doesn't require speed so much as power: super-wide buses, huge instruction sets, and the ability to send huge amounts of data around to various places pretty quickly.

That last part is the only part where "speed" as we generally conceive of it really comes in and it's generally dealt with by using superfast networking, because most supercomputers are just too large for information to be shuttled between components by pathways etched in silicone.

I suppose it may be easy to misunderstand this as well. Super computers are definitely fast. Faster than anything Alienware and $10k can get you. But most of their ability, the power, comes not from speed so much as that ability, previously stated, to do complex things with huge chunks of data all at once.

share|improve this answer
Interesting, huge instruction sets, as in CISC? I'd like to check out documentation of an example of a modern supercomputer with CISC's instruction set. –  Pete Oct 5 '11 at 18:06
I suppose there are computers that qualify as "supercomputers" that run x86, but I'd also imagine they're not that fast, relatively speaking. Other computers such as those used for weather prediction (Japan's got a few of these monsters) are most likely running completely custom architectures and instruction sets custom designed for running the complex modelling they excel at. –  music2myear Oct 5 '11 at 19:21
I'm not sure if CISC is a requirement for super or is a consequence of low costs. But the top500 statistics points that Power, Sparc and NEC processors just have 9.6% of the "market" there, with Intel and AMD having the remaining >90%. top500.org/stats/list/37/procfam –  woliveirajr Oct 5 '11 at 19:50
top500.org maintains a list so you can see what makes a "supercomputer" by their particular standards. There are a lot of big clusters of fairly work-a-day machines plus some special purpose architectures. –  dmckee Oct 5 '11 at 20:30
@music2myear: I think super computers should be more like powerful embedded systems. As they are designed to perform just one task, speed can also be optimized to a very large extend from the design. –  Fahad Uddin Oct 5 '11 at 22:55
show 1 more comment

no... supercomputers have regular memory, too. Example:

The Cray XT4, introduced in 2006 added support for DDR2 memory, newer dual-core and future quad-core Opteron processors.

source: wikipedia

Another example from Cray:

The Cray XK6 is an enhanced version of the Cray XE6 supercomputer, announced in May 2011.[1] The XK6 uses the same "blade" architecture of the XE6, with each XK6 blade comprising four compute "nodes". Each node consists of a 16-core AMD Opteron 6200 processor with 16 or 32 GB of DDR3 RAM and an Nvidia Tesla X2090 GPGPU with 6 GB of GDDR5 RAM, the two connected via PCI Express 2.0.[2] Two Gemini router ASICs are shared between the nodes on a blade, providing a 3-dimensional torus network topology between nodes.

source: wikipedia

share|improve this answer
add comment

By "cache technology" you are probably referring to the high-speed static RAM used for holding the memory data prefetched from main memory and/or not yet writtem out to main memory. You are probably not referring to the content-addressable memory used for holding the address tags.

There was a time (post ferrite core but before the 1990s and the proliferation of PCs) when static RAM was used to populate the main memory of computers. As demand for memory capacity (total amount of memory installed) and density (memory per square inch of board area) increased, the paradigm of a simple memory subsystem using SRAM evolved into using dynamic RAM (that needed refresh circuitry) with cache to compensate for slower memory read/write cycles. The huge shift from SRAM to DRAM production has made a small price differential into a huge one, in addition to the power consumption and physical density disadvantages.

A (super)computer could be built using SRAM rather than DRAM for main memory, but the cost-to-benefit ratio is low. SRAM only gets you raw memory speed, but DRAM augmented with modern cache technology can almost match that speed with less cost, volume and power consumption.

Memory speed is only one parameter in overall computer performance. For supercomputers processing throughput is often improved by using parallelism, rather than the fastest technology money can buy. Parallelism by doubling the width of memory data bus. Parallelism by using multiple processors. One of the first demonstrations of an inexpensive massive parallel computer was based on 64(?) Intel i386 processors at CalTech.

share|improve this answer
add comment

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.