OpenPOWER Summit 2020 | Hacker News Reader

OpenPOWER Summit 2020

phoronix.com

45 pointsprotomyth18 comments

I really want to like OpenPOWER. It has a lot of features I think are great, but it's just so expensive that I can't justify purchasing e.g. a Blackbird from Raptor CS to play with.

Is anyone aware of POWER-powered low cost SOCs akin to Arduino or raspberry pi?

Haven't heard of any readily available hardware, but here's a project you may want to keep an eye on:

https://www.powerpc-notebook.org/en/

It won't be as cheap as a Pinebook Pro. But even if it were in the circa $1000 range, I'd feel it would be far easier to justify buying than a Blackbird.

From https://static.sched.com/hosted_files/openpowerna2020/66/Ope...:

> We could not make use of commercial crowdfunding platforms:

> - they support the production of a product

> - they require a working prototype at the end

> We are not making or selling any physical product.

> We are doing and supporting research and development

In addition to this, even if this eventually leads to actual hardware, it will be based on a dead architecture. The PowerPC e6500 supports big endian ppc64, which is unfortunately a dead end. Since 2014 and the launch of POWER8, all Power development has gone towards little endian ppc64le. There will not be any ppc64 big endian only hardware. Running ppc64 Linux in 2020 (on e.g. a G5 Mac) can be buggy affair due to lack of attention towards Power or even big endian systems in general.

The QoriQ series is still the best choice for any Power-based laptop hardware mind you, it's just the state of affairs in Power that a 2013 chip aimed at network equipment is the best you can do.

What sort of problems are there specifically with current GNU/Linux distributions on big-endian systems? Fedora/RHEL, SUSE, and Debian are all supported on s390x, though doubtless many packages aren't very relevant on it. It was long ago that I ran SPARC Debian without issue that I remember, but it does seem still to be active.

I'm sure there is some badly-written stuff with issues, but I have the impression they're not widespread.

The most prominent problem you'll run into on BE systems is the lack of support of graphics cards post 2008-ish, and even cards that are supported tend to be very buggy. You can essentially say goodbye to 3D acceleration and hw video decoding, while ppc64le has good support up to the newest Radeon cards, and Nouveau works ... as good as it does on x86, anyway.

Package wise, you're also in a pretty miserable state. It's "usable", but even getting Firefox to compile on BE is a pain, and it's still a buggy experience. Firefox and Chromium build on LE systems, and with IBM's V8 port and patches to use Altivec and GPU acceleration, Chromium is very nice on LE.

Beyond those, more packages are broken on BE. As a practical example, there's the PPC void linux fork at https://repo.voidlinux-ppc.org/stats.html . Many applications do built, but there are a few blockers like the lack of Go support and lack of Haskell support, which trickles down into other packages (e.g. pandoc doesn't build on BE because lack of haskell). Neither docker nor podman work on BE.

Packages that do build are sometimes broken on BE as well. The more complex a program, the more likely it's subtly broken.

Basically, it's usable, but it's an extreme downgrade from LE, and requires being a lot more forgiving of your system. Enthusiasts using G5s are fine with it, but building a new system that's only BE capable would sting for eager customers receiving their inevitably pricey laptop only to discover a large number of prominent linux applications aren't available and the system is riddled with endianness bugs. Depending on what you use your computer for, you might not run into any problems, but you've got pretty serious problems across the "gaming, web browsing, productivity, devops" spectrum.

I actually meant the bit about "even big endian systems in general." I know of things missing on BE ppc64 from EPEL6 packaging. (There's a significant number of things of interest to me missing on ppc64le, but desktop stuff isn't the interest there.)

What I've seen most is issues with drivers. Even AMD, whose open GPU drivers make/made the planned included supplier of GPUs for this notebook have said that you're on your own on big endian systems. If you use the PCI Express slots on your 2005 Power Mac G5 to add a modern-ish (<12 years) AMD GPU to your big endian ppc64 system, it will give you a picture alright but 3D effects sometimes get corrupted and the hardware video decoder flat out does not work. The AMD driver simply implicitly assumes their platform is little endian in many places.

And that's not really a surprise. The only big endian system available to mere mortals that can fit an AMD GPU is a noisy, heavy, oversized metal box from 2005 you need to get off of Craigslist. Even the Raptor desktops default option is ppc64le. Most hardware developers never need to worry or care about big endian hardware, as none of their users use POWER, SPARC, S390X or something even more obscure.

This driver case is where AMD has made an explicit statement (though you'll have to forgive me from backing that up, can't find it anymore). In other cases it can be hard to differentiate between regular software bugs and endianness specific bugs.

Indeed. IBM could have done a lot more for Power than they are, but they don't understand how. They only understand how to make overpriced hardware for the enterprise. They don't understand how to build a grassroots developer community or the consumer friendly hardware needed for that. This has been a problem for nearly twenty years, as that problem is also the reason Apple switched away from Power.

Power and RISC V should look a lot more attractive because of the Nvidia/ARM deal, and Power has a lot of advantages over RISC V, with proven high performance designs and apparently really good Linux ecosystem support for ppc64le. But with IBM as the captain of Power, I highly doubt Power can capitalize on that at all.

>Indeed. IBM could have done a lot more for Power than they are, but they don't understand how. They only understand how to make overpriced hardware for the enterprise. They don't understand how to build a grassroots developer community or the consumer friendly hardware needed for that.

I'm inclined to agree and that's really disappointing.

There's a contrast between working with RISC-V and POWER covered at https://www.phoronix.com/scan.php?page=news_item&px=Libre-RI... which may be relevant, though I do rather agree about IBM, even while developing on a POWER9 system.

The relevant item from the summit is https://news.ycombinator.com/item?id=24535438 (sorry posted in parallel).

> POWER-powered low cost SOCs akin to Arduino

More in the range of the arduino than the pi, you could get a Colorlight 5A-75b for ~$15 on AliExpress and reprogram its FPGA to run an OpenPOWER core. It's a board meant for driving LED panels, but it has a Lattice ECP5-25F FPGA with a JTAG header easily accessible plus dual gigabit ethernet MACs and 4MB of SDRAM. The downside is all the rest of the IO is stuck behind one-way buffers since it's only designed to push data out to 5V LED panels. But it's been well reverse-engineered by the chubby75 project, the FPGA is well-supported by the open-source yosys+nextpnr FPGA toolchain, it's supported by the litex SoC builder (which also includes microwatt core support, though maybe not colorlight+microwatt?), and it looks like some have had success getting OpenPOWER cores to boot on it: https://twitter.com/antonblanchard/status/123177893988802969...

It will need some additional strategic investment from another decent size company. Currently it is still too tied with IBM on OpenPOWER CPU's production.

Unfortunately Broadcom, Qualcomm, Marvell, Apple, TI, etc all the big names in Fabless Semiconductor Industry haven't taken any interest. That might change with the Nvidia-ARM deal depending how certain companies management / engineers hate Nvidia. ( For what ever reason )

Personally I would love to have an A20 CPU POWERed Router to play around.

The problem with power 9, and likely 10, is that it was far too expensive and hard to get. Even two years after it was announced you had to talk to IBM salespeople, and the server was about 50% more expensive than an equivalent Intel. For a while they had pcie4 before anyone, and a strong backing from Nvidia. Those are both gone.

It would help legitimize openpower souch if someone would build some a20 based systems.

All eyes are on RISC-V right now, especially with ARM getting gobbled, but the game is open to anyone, any architecture, & it all comes down to making good chips at good power & price points.

Skip the hardware slide, it's mostly cringeworthy marketing fluff with little solid technical detail.

best-of-breed/Enterprise Focus/Powerful Core = Enterprise Strength + AI Infused/World-class/Enterprise Cloud/Powerful Core : Enterprise Strength/you get the picture

edit: take that back a little, the 2nd half gets a bit hardcore

Power is an irrelevance now.

Reply on news.ycombinator.com