Hacker News Reader

Comment by artemonster | Hacker News Reader

as a HW designer that writes RTL for living I will repeat this 150 times and will put this on my gravestone: WE DONT NEED ANOTHER SHMANCY HDL. really. existing ones are moooooorrreeee than fine. our tools suck, verification sucks. your design complexity is entirely limited by your verification capabilities (and automation infra). having fancy constructs for CDC or pipelining in HDL is utterly useless, especially that CDC checking is done by special tools that do it nearly perfect with a bit of constraints.

View on HN

As a hardware engineer turned software developer (including in EDA) who has dabbled in trying to create higher-level HDLs over the years, I now tend to agree with you. As it turns out you can already do a lot of things with parametarization of modules (which has been possible in Verilog and VHDL for years (decades) now). I think a lot of SW folks who look at the problem tend to come up with something that seems better to them from a SW engineering standpoint, but also tends to ignore some of the special needs of HW design so it ends up being klunkier than just writing VHDL or (System)Verilog.

When Javascript developers don't understand something, they "replace" it because NIH and failure to understand the subject mater. They have hammers, and so everything is a nail to be replaced.

Totally agreed. This is the problem of academia unfortunately, the people working on these have no experience in designing complex chips and facing the real limitations.

We are so stupidly limited by our EDA tooling and infrastructure. I wish these efforts would have been put to use in that front.

It would be nice to have a simulator like Verilator for VHDL.

There is mature GHDL https://github.com/ghdl/ghdl and rather new NVC https://github.com/nickg/nvc.

I was aware of GHDL. NVC looks like it's potentially more performant.

LLVM never works in EDA.

> We are so stupidly limited by our EDA tooling and infrastructure.

The problem that no one will ever solve is there's no gcc equivalent to NXT. Everything is downstream of that problem.

are you sure it will do anything at all? I am not. The current setup, while objectively sucks for engineers is still quite capable and it *works*. we are far from reaching limits of what is capable to design and tape with current flows. frontend money aspect for tools is still peanuts in comparison to backend and actual fabrication, so it will not generate a renaissance era like gcc did for software

Backend and fab costs are dominant (or even packaging costs are on par with silicon costs these days), correct, but verification is an multiplier on overall costs. The cost of a re-spin is huge. To your point though, most issues warranting a resping are backend related.

> are you sure it will do anything at all? I am not. The current setup, while objectively sucks for engineers is still quite capable and it works

go back in time to before gcc - i'm 100% sure people were saying the exact same thing about borland (or whatever).

I don't recall anyone being "objectively-sucks" level of unhappy with proprietary C compilers. Moving to gcc was often a regression -- because it was much slower to run and sometimes generated worse code -- for developers used to Turbo C or CodeWarrior-style IDEs.

AFAIK when it comes to using the features of recent CPU architectures, Intel's compilers produce more performant / efficient code than GCC or Clang.

A sufficiently smart compiler requires ever more clever compiler writers who are deeply knowledgeable about the many quirks of the numerous architectures.

> Intel's compilers produce more performant / efficient code than GCC or Clang

Intel's compiler are clang forks.

Word. 28 years of FPGA and ASIC design here, in VHDL, Verliog and SystemVerliog. Coming from VHDL, verilog had some painful limitations (no struct/record type) but SV fixed those, and supports some surprisingly powerful metaprogramming.

But even when using plain verilog the language was never the limiting factor on the design process.

Reply on news.ycombinator.com