bcm5719-fw VS ompi

I care! And I know a lot of people who care, but we are still a niche sized group. I care mainly because of Raptor Computing Systems offerings, which I think are the main (only?) OpenPOWER systems available. I use a Blackbird, and I'm happy with it.

From my own point of view, I'm willing to spend a $$$$ premium on hardware where I can have assurances that from the time I boot it, only code I authorize to run is run. Where every part of the system has code that, at least in principle, I or someone else could audit and fix. People have valuable IP stored on computers and it's worth much more than a few thousand dollars.

If you just look at price to performance, you are missing the point. Also, the price is not out of line with other niche desktops such as Apple's or System76.

There's not a lot of competition in this niche. The previous system that was useful was a ASUS KGPE-D16 motherboard, which could be librebooted (https://libreboot.org/docs/hardware/kgpe-d16.html) I expect something new to come along in this space every 5-10 years.

For my purposes, I haven't fought with the software ecosystem, and was able to compile the very few packages that weren't already precompiled.

Here are some developments I think are worth noting:

* There is a libre driver for the onboard NIC. (https://github.com/meklort/bcm5719-fw) This seems to be the only project that cares about blobs in every part of the board.

* Dasharo https://www.dasharo.com/ providing alternative boot firmware.

* Artic Tern, (https://www.raptorcs.com/content/AT1PC2/intro.html) which is objectively still mostly a development platform (that if you're skilled you can get to work) provides a completely libre boot environment and the possibility of controlling other peripherals using only auditable code.

A few things have not yet made it onto the board:

* Flexver (https://www.raptorengineering.com/TALOS/documentation/flexve...) which would allow for verifying and auditing hardware, firmware and the boot process isn't commercially available yet.

* Ultravisor state enabling more secure VMs is still awaiting implementation AFAIK. (https://wiki.raptorcs.com/wiki/Power_ISA/Privilege_States#Ul...)

* I'm not aware of a lot of hardware that would take advantage of IBM CAPI 2.0 IO accelleration. Perhaps someone has some information on this.

* I'm not sure what the status of transactional memory is, but I'm not aware of it being used in software. Perhaps someone can enlighten me on this.

These would be nice to have, and I hope to have them in the future.

The bottom line is that this is the only hardware currently in production that is going in the direction promised by the personal computing revolution back in the 1970s and 80s and is still capable of handling most people's current general computing needs. I write this hoping that other people like me who are reading this understand the importance of keeping hardware like this alive.

Here is the GitHub repository: https://github.com/meklort/bcm5719-fw

bcm5719-fw - an alternative firmware for the network card Broadcom BCM5719;

bcm5719-fw - I saw announcement of this project on Raptor Forum. I will ask melrott, if there is intrest in introducing project at vPub.

The Talos II is blob-free. At launch, proprietary binary-only firmware was required for the network interface, but Raptor Computing Systems offered a bounty to reverse engineer and do a Free Software re-implementation of the firmware, and that effort succeeded and the bounty was paid. See:

https://wiki.raptorcs.com/wiki/Project_Ortega

https://github.com/meklort/bcm5719-fw

A clean-room RE requires one party to determine the behaviour of the original product, and write a spec for it. Then a second, completely different party must build the reimplementation using only that spec. They can't communicate with the first party through any means other than that spec. For a pretty clear example of this practice at work in open source, see the work to reverse-engineer and then reimpliment an open-source driver for the Broadcom BCM5719 NIC.

Adding to your great list, https://github.com/meklort/bcm5719-fw/ . Clean-room firmware for the onboard Broadcom NICs in the Talos II / Blackbird.

The commercial Linux world (see https://github.com/open-mpi/ompi/issues/4349) and other open source OSes (eg FreeBSD) seem to have lined up behind little-endian PowerPC. IBM still has a big-endian problem with AIX, IBM i, and Linux on Z.

Roughly, the sets of computational problems that people used (use?) MPI for. Things like numerical solvers for sparse matrices that are so big that you need to split them across your entire cluster. These still require a lot of node-to-node communication, and on top of it, the pattern is dependent on each problem (so easy solutions like map-reduce are effectively out). See eg https://www.open-mpi.org/, and https://courses.csail.mit.edu/18.337/2005/book/Lecture_08-Do... for the prototypical use case.

One of the guys at the local LUG is one of the lead maintainers of Open MPI. He told us about a comment that ran into the hundreds of lines, all for a one-line change in the code.

But it would be very inconvenient to have to keep crediting everyone who's ever worked on it. If you look at old projects, their licenses can have like 10-20 of those lines (here's one I was recently looking into).

I have a bit of experience programming for a highly-parallel supercomputer, specifically in my case an IBM BlueGene/Q. In that case, the answer is a lot of message passing (we used Open MPI [0]). Since the nodes are discrete and don't have any shared memory, you end up with something kinda reminiscent of the actor model as popularized by Erlang and co -- but in C for number-crunching performance.

That said, each of the nodes is itself composed of multiple cores with shared memory. So in cases where you really want to grind out performance, you actually end up using message passing to divvy up chunks of work, and then use classic pthreads to parallelize things further, with lower latency.

Debugging is a bit of a nightmare, though, since some bugs inevitably only come up once you have a large number of nodes running the algorithm in parallel. But you'll probably be in a mainframe-style time-sharing setup, so you may have to wait hours or more to rerun things.

This applies less to some of the newer supercomputers, which are more or less clusters of GPUs instead of clusters of CPUs. I imagine there's some commonality, but I haven't worked with any of them so I can't really say.

[0] https://www.open-mpi.org/

I would suggest using OpenMPI because it's pretty easy to get started with. You can build OpenMPI with CUDA support, then you can pass device pointers directly to MPI_Send and MPI_Recv. Then you don't have to deal with transfers and synchronization issues.

backends from native torch distributed configuration: nccl, gloo, mpi.

I use a linux ubuntu machine with MPI (https://www.open-mpi.org/). I had a question on making my computer simulations faster. Would be better to get an older AMD 9590 machine clocked at 4.7 ghz or continue using my Ryzen 7 1700 machine clocked at something like 3.5ghz?

OpenMPI - Message passing interface implementation. BSD-3-Clause