News for hackers. Stuff that matters.
Re:ARM or x86? ISA Doesn't Matter by remexre on Saturday August 31, 2024 @07:47PM (#41412927)
> it has got the right top eight instructions that you actually need to optimize for

I wonder which ones he considers those to be -- the interview doesn't say, sadly...

Re:ARM or x86? ISA Doesn't Matter by janice1999 on Saturday August 31, 2024 @07:54PM (#41412964)
They've made some details public about their architecture [1] and even their RISC-V test suite [2] but I'd also love to know what it was referring to.

[1] https://www.theregister.com/2024/08/27/tenstorrent_ai_blackh...

[2] https://github.com/tenstorrent/riscv_arch_tests

Re:ARM or x86? ISA Doesn't Matter by anonymoushn on Saturday August 31, 2024 @07:53PM (#41412959)
I'm hoping we at least get pshufb, aesenc, rep movsb, some sort of integer multiply, and a floating point fma
Re:ARM or x86? ISA Doesn't Matter by rowanG077 on Saturday August 31, 2024 @07:56PM (#41412969)
I would say it's load, store, compare, add, subtract, call, return and branch. If you have that you cover almost all code.
Re:ARM or x86? ISA Doesn't Matter by sgerenser on Saturday August 31, 2024 @09:34PM (#41413499)
Implementing multiplies as a bunch of adds has got to suck for performance though.
Re:ARM or x86? ISA Doesn't Matter by colejohnson66 on Saturday August 31, 2024 @10:25PM (#41413714)
Well, if you decompose it to a series of shifts and adds, it’s a bit quicker. That’s actually how the 8086 implemented it.
Re:ARM or x86? ISA Doesn't Matter by rowanG077 on Saturday August 31, 2024 @11:13PM (#41413961)
Sure. Multiplication is exceedingly rare compared to any of those 8 instructions though. RISC-V base instruction set doesn't include multiplication for example. If you had room to add more then 8 I would rather add the bitops first. This will allow you to also write a much more efficient soft mul in addition to many more things.