3 ms·
Supercomputers don't so much buy speed as time. A supercomputer is a decade or two ahead of consumer-grade hardware of the time. And it's not just a matter of
by dredmorbius 5y ago
Supercomputers don't so much buy speed as time.
A supercomputer is a decade or two ahead of consumer-grade hardware of the time.
And it's not just a matter of processor speed. Parallel disk I/O, total memory size, cache, and a number of other tricks factor in.
Today's cluster supercomputers are mostly just ("just") racks and racks and racks of fairly standard-grade CPUs, with high-speed switches.
GPUs have introduced their own curve into things. They work well for highly-parallelisable problems. Unfortunately, the ones that seem to have caught many people's attention are ultimately somewhat boring.
- agumonkey 5y agothere's a thread on slashdot IIRC where a guy explains that a good old cray, due to its massive cache and bandwidth, can crunch number matrices at massive "speed". I forgot the details but basically each cycle the processor could recompute a whole 1MB or 8MB matrix, and steadily so. Whereas a pentium4/core cpu, clocked way higher and more "modern" can only process that amount of data for a small time before all caches are filled and performance plummets.
- bombcar 5y agoPeople compare supercomputers to general purpose computers and forget things like this - the supercomputer and the program it ran were often designed in parallel to do a particular thing very very fast.
- agumonkey 5y agoYeah goals impacts architecture. But it's interesting to see that computing systems have parameters and structures. Cycle clock, memory, cache .. they don't paint the whole picture. The constant massive data crunching of a Cray is not visible on frequency. That said, I'm curious about what a 2020 iphone can do, even in HPC numerical contexts..