What is a “super computer” and what can it do?
In: Technology
Super computers are purpose built to perform computing tasks at a higher level of performance than a commodity, retail, or workstation computer. Your phone today is several orders of magnitude more powerful than the first super computers, such as the famous Cray-1, but that doesn’t, in some way, demote a super computer.
Typically, super computers use processors that are designed for math operations, very much like a video card’s GPU. There is an encoding into binary bits of “real” numbers called a “floating point”, and these computers perform many FLoating point arithmetic calculations. The number of **FL**oating point **OP**erations a computer can perform per second is called a **FLOP**.
There are, or were, general purpose super computers that weren’t especially focused on floating point operations, and they measured the number of CPU instructions they could perform per second, not just floating point instructions, and they measure their performance in **M**illions of **I**nstructions **P**er **S**econd, or MIPS.
What they can do is kind of hard to say beyond the above, without also talking about what they’re used for. Today, most super computers are used for physics simulations. These simulations are most often for weather, but also for nuclear physics, astrophysics, material sciences, chemistry, and engineering.
There is a long history of super computing. The earliest machines were analogous to superscalar single core machines, compared to their contemporaries. These early super computers pushed the computational boundaries of their days regarding materials, signaling, and propagation among the circuitry of the day. Cray-1 is regarded as one of if not THE first super computer, and while integrated circuits were invented by this time, they were so new, Seymour Cray basically didn’t trust them? He built his first machine using discreet components.
Eventually, miniaturization smacked face first into the physical limits of quantum physics, so smaller and faster was no longer a way to chase down performance. We call this vertical scaling. So the machines stopped scaling vertically, and instead went horizontally. More cores, more nodes.
If you saw a super computer today, they would be racks and racks of nodes in a data center. These are not individual computers in a closed network! These nodes cannot “boot” and operate on their own, they don’t each run their own operating system, they don’t each have their own video port, or their own disk drives. These nodes are little more than CPUs, and do not operate in a standalone fashion. The network that combines the nodes aren’t your typical Ethernet network, either. The sum of these nodes and the network that interconnects them constitute a single computer with hundreds of thousands of cores. Your computer has many compute cores within it – if you just look at task manager. These computers have many cores within each CPU, and many CPUs, not just within a single node but spread out among hundreds and thousands of nodes.
Then there are compute clusters. These are a distinct class of computing. They’re not super computers, though they will get called that often enough. These are indeed individual computers, each one can run standalone, each one made from off the shelf parts. They’re just arranged with a private network and special software to sort of treat each node in the compute cluster as a worker node. Work is farmed out in batches to nodes that need work. You can assemble one of these in your own home, and they’re popular among businesses and universities on a budget. A classic cluster computer setup would be a Beowulf cluster, though not relevant in the modern era.
There are grid computers. This is similar to, but distinct from a cluster in that each unit in the grid can do different work, whereas a cluster is focused uniformly on a singular goal.
Then there is distributed computing. This is where computers all across the internet can download a client that helps manage the work to be done. SETI@Home and Folding@Home were and are popular, where that since your computer spends most of its time idling, it uses that idle time to perform work.
And then there are mainframe computers. These are not super computers, and the term does not mean some really old computer of a bygone era. In fact, the business behind mainframe computers are alive and very well, won’t go away any time soon, and is even expanding. Mainframe computers do things that none of these other computers can do, and that is they perform transactional computation – an all or nothing sort of thing; either the computation is complete, verified, and committed, or not at all. This is *exactly* the sort of computation you want when handling financial transactions, like when you pay with a credit card or transfer money. The other aspect of a mainframe that unique is the high throughput, which is a very specific metric. Just one of these mainframe computers can perform billions and billions of transactional computations in a single day, non-stop, uninterrupted, for years or decades at a time. I once got a tour of one of these machines, which takes up several racks in a data warehouse, and the largest ones made today have a couple thousand processors and modules just dedicated to moving and managing data to keep the CPU fully saturated with work. Nearly the entire financial system worldwide is built upon these things, as are many government systems. When COVID hit and the US unemployment office got completely slammed, remember when New York Governor Cuomo blamed the decades old mainframes? Nope, those babies were humming along just fine, underwhelmed. It was the more conventional web frontends that absolutely crumbled under the load. But mainframes are not super computers, they are only good for the type of work that they do and are thus very special purpose.
There is cloud computing, which is what my company does, and that is basically paying someone else to manage the hardware, and the client rents use of computer resources. Lots of companies choose cloud platform providers because owning and maintaining hardware can be expensive. Some providers have performance oriented offerings.
Finally, there’s High Performance Computing. This is like a mini-super computer. You see the Bitcoin miners do stuff like this all the time, where they’ll either stuff a computer with as many video cards as possible for their math processors, or they’ll use custom FPGAs or ASICs (basically, purpose built custom processors).
Latest Answers