5 comments

  • bem94 4 days ago

    It's great that companies do this, but I wish that they would also open-source the verification infrastructure at the same time.

    Otherwise, anyone who wants to use it in a commercial setting has to either re-do their own verification effort and due-diligence, or just trust the core works perfectly. Don't get me wrong, I trust IBM to build a working CPU, but that doesn't mean there aren't corner cases to be found when you put the core in a new SoC environment.

    There is lots of open source HDL out there. We need re-usable verification infrastructure too!

    • orbifold 4 days ago

      Most of that verification infrastructure is tied to proprietary tooling in some way or another. In the case of IBM they have a lot of in house EDA tooling that they aren't willing to release or even talk about much. They extended vhdl with "aspect oriented programming" to implement their approach to verification. Most of the other verification tooling for verilog (now system verilog) started as proprietary extensions by some vendor (Intel, Synopsis, ...) and then made it into the system verilog standard. None of those verification parts of system verilog have an open source compiler that supports them.

      The good news is that at least the simulation part has improved massively over the last few years. For synthesizable system verilog you have https://www.veripool.org/wiki/verilator, which allows you to compile to C++ and then you can implement any verification methodology you like from scratch in C++ :). I did that for parts of an in-house POWER based processor and it isn't actually as bad as it sounds. You also have the option to use DPI-C and interface with a C++ test-driver that way. We managed to do extensive Hardware and Software (full hardware abstraction layer + instruction compiler + FPGA + ASIC) co-simulation for an accelerator we build, although using proprietary tools because we had Xilinx IP and VHDL involved. Verilator is also used as a backend for Chisel.

      For vhdl there is https://github.com/ghdl/ghdl.

      • rbanffy 4 days ago

        > which allows you to compile to C++

        Sometimes I wish there was a tool that'd allow to compile MAME code to hardware and VHDL to code that'd be pluggable into the rest of MAME.

        That'd be something.

        • monocasa 4 days ago

          Higan (née bsnes) is written in a very HDL way. It's almost entirely an engine for evaluating RTL graphs with coroutines. cen64 is very similar. 'On clock edge, evaluate the next clock's input latches'.

          But yes, it would be really nice to increase verilator's perf to be close to higan (which is a multi dimensional problem, latency versus throughput, etc.), so that you could compile to software or hardware. A MAME cabinet based around an FPGA with cached configuration images would be the bee's knees.

      • Exactly and note "The A2I core is compliant to Power ISA 2.06 and will need updates to be compliant with either version 3.0c or 3.1. Power ISA 3.0c and 3.1 are the two Power ISA versions contributed to OpenPOWER Foundation by IBM."

        Who is going to do that and how without the verification infrastructure? (Verification is about the test infrastructure, not the Verilog simulator which is a commodity). The _real_ value is in the verification infrastructure.

        • aarong11 4 days ago

          This is probably high up on the list of reasons why they don't.

        • raxxorrax 4 days ago

          This should be interesting. Probably far too complicated for me to understand, but will still take a look. Surprised to see VHDL instead of Verilog, but not complaining at all.

          edit: Just took a look. Code is nice and clean but complicated was an understatement.

          • monocasa 4 days ago

            Yeah, IBM has been all in on VHDL for decades.

            And complexity-wise it's a lot better than quite a few HDL codebases I've seen, just very modular, and probably missing some high level design docs to ease getting your feet wet.

          • monocasa 4 days ago

            > It is a 27 FO4 implementation,

            Am I out of touch, or is that an extremely long cycle? I thought CPUs were generally something like 14-15 FO4.

            • pjmlp 4 days ago

              Composed of VHDL and Tcl source code, nice to see them pursuing the right path. :)

              • olympusultra 4 days ago

                Good news! But will it be sufficient to reverse declining popularity of the POWER architecture?

                • rbanffy 4 days ago

                  IBM targeted POWER towards the high-end server market. While this space suffers erosion from x86's coming from its low-end, IBM is not known to play in the commodity price-pressured segments unless it can't avoid it. I'm very sure that if IBM wanted to compete in the HEDT segment against Intel, it'd be able to offer a competitive product (as the Raptor workstation proves).