[dmd]

> way slower than your phone

For the young'uns... I think you're underselling this a bit. A really powerful late 70s mainframe, generously would have been somewhere around 1 to 2 million instructions per second.

Your phone is probably on the order of 5-10 trillion instructions per second. That's a million times faster.

[brk]

True, but 90% of the mainframe instructions weren't tracking, telematics, data exfiltration, and user profiling. So it kinda evens out.

[The_Colonel]

> Your phone is probably on the order of 5-10 trillion instructions per second. That's a million times faster.

Do you mean from iGPU? The fastest smartphone iGPU I found is MediaTek 9300 which has 2.4 TFLOPS which roughly corresponds to your claim. Cray-1 (1975) had 160 MFLOPS which would make today's high-end smartphone 15 000 times faster than Cray-1.

[Jtsummers]

Cray-1 wasn't a mainframe, it was a supercomputer. As such it was designed to push the performance envelope in a way mainframes (of the time, but even today) weren't even trying for. Mainframes were designed for transaction processing and reliability. They were substantially slower than contemporary supercomputers.

[The_Colonel]

That's true, but supercomputers are somewhat conceptually similar to GPUs, so it seems like a better comparison to illustrate technological progress. Comparing mainframes and GPUs is comparing apples and oranges.

[zifpanachr23]

This is a good analogy. If you look at a modern mainframe CPU it becomes pretty clear where the differences lie. Fewer, beefier cores with a lot of focus on cache.

Supercomputers tend to use more conventional cores, but way more of them, and connect them in a large fabric. There's a lot more focus on parallelization and horizontal scaling.

Mainframe overall compute is nowhere near a supercomputer, and you probably shouldn't be running a massive physics simulation on a mainframe, but you may get more consistency and reliability for well defined tasks.

[vaxman]

This table will help you, but it stops way back in 2021 at 346,350 MIPS https://en.wikipedia.org/wiki/Instructions_per_second

Looking forward to the Apple A18 Pro reported to be 4GHz x2 and 2GHz x2 +/- thermal conditions.

[Taniwha]

Yes exactly, core cycle times were ~1uS so around 1Mhz clock for our old B6700

[kragen]

your phone is 5–10 billion instructions per second, a thousand times faster. it's probably a quad-core in-order arm running at 1–2 gigahertz, not 1–2 terahertz

[dmd]

Instructions and hertz have not been equivalent for almost a quarter of a century now, though.

[kragen]

not equivalent, but close enough in this case; even my laptop only manages less than 2 ipc per core, and the error in your earlier comment is a factor of 1000

[vaxman]

I think we need to put it in terms of smart watches now, soon smart rings.

Hope we remaining masters can stop the delusional tech bastards running around convinced they're a new and more powerful Steve Jobs (oblivious that his success came from leveraging microprocessors to put millions of formally educated tech people with families out of work in a disorderly manner by disrupting the market) from further inflating the public stock markets (again) and distorting energy capX over AI hype. AWernerS is clearly losing the battle against them. https://www.youtube.com/watch?v=ixe8Snxu3wo