21 ms·

Early software was designed to feel as fast as possible, even resorting to hacks like temporarily turning off DRAM refresh to write text on the screen as quickl

Early software was designed to feel as fast as possible, even resorting to hacks like temporarily turning off DRAM refresh to write text on the screen as quickly as possible. Though hardware was very limited, when you pressed a key you generally got a response right away. Fast forward to 2023, and everything runs on top of a non-realtime multitasking OS, causing periodic little glitches as background stuff hogs up resources. Graphics is too complicated to guarantee consistent 60 frames per second. And web is incrementally loaded from shared servers, often with delays of many seconds. Animations introduce intentional lag. Modern hardware could be of course amazing at being fast, but nobody put in the effort into software designed for that goal. Plus obsession with device thinness and fanless designs introduces thermal throttling that no software can overcome.

ElectricalUnion 3y ago
> Modern hardware could be of course amazing at being fast, but nobody put in the effort into software designed for that goal. A lot of effort has being thrown that way, it's just that their definition of "go fast" is "having lots of throughput", not the things you are looking for: goodput, low latency, low jitter. We got faster mainframes instead of faster minicomputers - computers and network systems that are optimized at doing batch jobs. We can submit a whole bunch of blocks and the graphics processing unit can display accelerated smooth video for us. Or we can push a whole neural network to a tensor processing unit and have it do inference in very few operations, after the model is loaded. But both of those operations while having smooth output have horrible startup latency. I think is very naive to call what the devices have today as a "single computer" when in fact, for a long while they're several interconnected computer components joined with lots of buffer.

ElectricalUnion 3y ago

> Modern hardware could be of course amazing at being fast, but nobody put in the effort into software designed for that goal. A lot of effort has being thrown that way, it's just that their definition of "go fast" is "having lots of throughput", not the things you are looking for: goodput, low latency, low jitter. We got faster mainframes instead of faster minicomputers - computers and network systems that are optimized at doing batch jobs. We can submit a whole bunch of blocks and the graphics processing unit can display accelerated smooth video for us. Or we can push a whole neural network to a tensor processing unit and have it do inference in very few operations, after the model is loaded. But both of those operations while having smooth output have horrible startup latency. I think is very naive to call what the devices have today as a "single computer" when in fact, for a long while they're several interconnected computer components joined with lots of buffer.