Hacker News new | ask | show | jobs
by PaulHoule 1545 days ago
What ramblings.

Optane is the best performing SSD but the worst performing RAM you ever had. It is too expensive at any speed, even if Intel is losing money on it. HP memristors are vaporware.

LISP machines, Java machines, and similar architectures specialized for complex language runtimes are a notorious dead end. They just can’t keep up with performance-optimized RISC, pipelined, superscalar, SIMD, etc. architectures paired with compilers and runtimes that implement efficient abstractions (e.g. garbage collection, hotspot compilers) on top of those very fast primitives.

4 comments

Before Lisp Machines were killed in the market it was clear that new architectures were needed and a few were under development, even RISC like CPUs. They weren't released.

But Lisp at that time was already fast enough on standard RISC chips (MIPS, SPARC, ALPHA, POWER, ...). Later the 64bit RISC chips also provided enough memory space. SPARC also had some tricks for Lisp implementors.

Currently the assembler coded Ivory emulator is 80 times faster on Apple's M1 than the last Ivory hardware (the Ivory Microprocessor from Symbolics was released end 80s).

Speed is relevant for some use cases, sure, but not at all for a ton of others. Memory, disk and CPU are almost free in this new world, so why are we computing like it's 1990 still? It's time for some different abstractions than file -> process -> file.

The vast productivity gains of Smalltalk and Lisp were because they discarded those abstractions and programmers were free for others.

Presumably OP posted this after noticing Phantom came up a few days ago. https://news.ycombinator.com/item?id=30807668

> Memory, disk and CPU are almost free in this new world, so why are we computing like it's 1990 still?

Elsewhere on this very site you'll find no ends of complaints about, say, Electron apps.

For general purpose computing applications expand to fill the performance available (that includes real value and bloat!)

I dabble in microcontrollers for fun and there it's different. I am an AVR-8 fanatic and sometimes I think "this is so fast" and "2K of RAM is plenty" and "I can fit CRC-32 tables in 32k of flash because that's what counts as an 'operating system' for me"

Then there are the applications where it just doesn't have the power and I am so glad to have a box of RP2040's because in 2022 the most important attribute of a microcontroller is that it is available.

The RISC-V folks are working on additions for special support of "complex language runtimes". Pipelined, SIMD and superscalar are all well and good, but what kills pure software-side support is always heavy branching and dispatching. These operations are genuinely much faster and more power-efficient when implemented in hardware.
How is the ARM not a "JavaScript Machine"?

https://stackoverflow.com/questions/50966676/why-do-arm-chip...

>Why do ARM chips have an instruction with Javascript in the name (FJCVTZS)?

https://community.arm.com/arm-community-blogs/b/architecture...

That instruction is a very small hack that uses just a few transistors to speed up a bit of data conversion that JS runtimes do frequently. That’s a far cry from a specialized chip.