Heat, not cleverness, now sets the limit on how fast computers get
For nearly two decades, from 1986 to 2003, computer hardware got more than 50 percent faster every year. Then the gains slowed. The binding constraint became getting power into chips and heat out of them, which is why modern processors quietly throttle themselves and data centres bathe machines in cooling fluid.
Hardware is the physical side of computing: processor, memory, motherboard, storage, graphics card and case, plus screens, keyboards and speakers. The name reflects rigidity. Software is soft because it is easy to change; the machinery is not. Its lineage runs through Blaise Pascal's adding machine of 1642, Gottfried Leibniz's stepped reckoner of 1676, which could multiply and divide, and Charles Babbage's never-built general-purpose engine, which already had punched-card input, memory and an arithmetic unit. George Boole's true-or-false algebra, devised in the mid-19th century, now underlies every chip's circuits.
In 1945 John von Neumann, working on the ENIAC project, outlined the template most computers still follow: one memory holding both data and programs, a central processor and input-output units. Because data and instructions share a single pathway, fetching both at once creates a traffic jam known as the von Neumann bottleneck. Later came reduced instruction sets in the 1980s. RISC designs, keeping only the commonly used commands and adding pipelining and caches, displaced more complex designs wherever power or space was tight, as in mobile phones.
With single-core speed gains fading, performance now comes from parallelism: doing the same operation across lots of data, as graphics processors do, or running separate tasks at the same time. Memory is arranged in tiers, with small, fast, expensive memory nearest the processor and big, cheap storage farther out.
Chips are usually more powerful than their coolers can sustain. They can briefly run faster while cool, but then slow down or shut off to protect themselves. Air cooling is common in smaller data centres, while immersion and direct-to-chip liquid systems cost more but work better. Machines range from phones built around battery life to room-filling mainframes and supercomputers costing hundreds of millions of dollars, whose internal networks matter as much as their processors.
Source: Computer hardware
