This has mushroomed and I'm so glad to have the opportunity to reveal the entire mess here: The Alpha XP900's PCI cage is no different than any 90's cage. It uses PnP, it accepts and assigns cards intelligently and properly. it has no discrete tool, like a 90's PC, but it is the same. If you get it confused with a stupid cage-stuff: Take all the cards out. Stuff the cage with RTL nics or VIA usb cards, neutral, friendly, the bus will love them. If it makes you feel better, boot bsd and run them on the bus. There's no way the cage remembers anything after that. This problem never happens, but there's your fix. To propery orient the machine, prefetch_mode should be off for the first probe by the booted OS, even if it crashes your machine. Then turn it back on and boot again. This clinches the PCI cage and the EV6 bus through Pchip etc.... The problem open-source developers had was they could not be certain of the hardware config. I can be. They were bouncine their code off of unknown hardware and, having nothing to hang their hat on, eventually quit, blaming the hardware. Each time they ran a squirrely, unfenced driver, they skewed the bus for everything. It was untestable. I think I destroyed a (very cheap) Si3114ctu controller chip with netbsd's satalink driver. I got 80meg@sec (lucky boot) and beat the hell out of it, then it up and died.... I think all the extra traffic and work involved literally worked the chip to death. Add this to the mix as a possibility and it re-enforces in the mind of userland the machine is junk. So the ultimate answer is that the Open-Source Community, userland specifically, screwed the pooch for every weak-memory-model architecture they ever ported to, by ignoring page one of the Alpha architectural white-paper. NetBSD, OpenBSD, FreeBSD, and Linux kernels always booted because they were written for the Alpha and obeyed the need for memory barriers to enforce in-order messages to the card-cage. All userland But Drivers use kernel for thier work and are thus protected by (netbsd's) Bus_space(), or other similar mechanism. The Tsunami does't give a hoot what order things come to it, but PCI is linear and it sure does. Since the kernels were all compiled with memory barriers, why doesn't X run on any of them? Why does the onboard ata controllers on the Alpha machines universally throw orphan (stray) irq's and downgrade to PIO mode 4 after initial negotiation at UDMA2? Userspace is not compiled with memory barriers at all. Xorg just pukes out of order garbage at the pci cage and stuns the video card, crashes the entire bus as often as not. This is All the BSD's and All the Linux. It is vital that these folks do something, for Risc V has a weak memory model. The Alpha is by far not alone: The PowerPC FX970 also suffers from this, but the iMac G5 is a very tightly integrate machine with no expandability and so survives the bus-storms, but they are there in userspace, sure as God made little green apples. I don't think it helps performance at all on PARisc.... All of these would benefit from increased stability and better disk and network IO! if userspace were properly curated. The Alpha was just the First victim of a blind-spot so enormous it is hard for me to credit, but every time I look at a release, I see the same sin, over and over and over. No Bus_space(), no Alpha_mb() ANYWHERE but the kernel(s). Some new discoveries in-the-bud: I feel very strongly that prefetch_mode in the srm does Not require INIT to toggle. I think it is toggleable from there as often as you like, in fact I think it is toggleable at a speed measured in at least the kilohertz.... I think this is the toggle, with another one or two, maybe, which NT and OpenVMS and Tru64 are able to achieve such good performance on a bus which requires in-order data and is hanging in a bag off of the Tsunami bus.... I think that these three operating systems have discrete rules they enforce using these toggles. We don't know where they are, but prefetch is one we do and perhaps can probe for in memory....? Even if this dynamic mechanism is unreal ( I think it is real ), or even if we can't exploit it (very likely, especially in short-term), it doesn't really matter. We've always accepted Some penalty for open- source and with the proper build it will run very very well with no need to resort to the above speculated optimizations. So, in short, the Alpha is clean. The hardware is good. Modern hardware worke Better than period because it knows PnP better, but it all works unless you are just a mad bomber. All the BSD's and linux's work. NetBSD 9.2 works best by far on this machine. I had X running for a week on a lucky boot with a clean cage and a following wind. Heavy compiling all week (over a very quiet and well-behaved Intel 1000 pro lan card). I'm not a coder so the software side is up-hill, but I'll get there. My first goal is to try and patch the wd0 driver for memory barriers. I expect it will instantly negotiate to PIO4 UDMA2 and stay there without any 'stray interrupt 14, stray interrupt 15' garbage in console. That will validate this entire sad episode and I'll do it first if someone doesn't beat me to it. Please dig out your Alphas. There was nothing wrong with them to begin with. On my Facebook post I inluded some photos of a pirate copy of High- Performance Computer Architecture, by Harold Stone. It is a clean-room copy, a re-set, and there is only one in the world. I would love to show this to Mr. Stone and tell him how it came to be. Here's a link to a couple of pages of it: https://www.facebook.com/groups/retrocomputers/posts/28340501522219693 Thanks for listening, thanks for being here all these years. I didn't realize until this other day, but the list started about the time I joined it, around 1998/2000. Then we were just a bunch of geeks, but today classicmp is an institution, venerable and respected. BZ! . BZ!On Tue, 2026-07-07 at 12:00 -0500, cctalk-request(a)classiccmp.org wrote: 0500, cctalk-request(a)classiccmp.org wrote: