Refurb weekend: DEC AlphaPC 164LX
oldvcr.blogspot.com
oldvcr.blogspot.com
I started out as a sysadmin for an stats department in 1993 that made heavy use of DEC Alphas running DEC OSF/1 due to their floating point performance. DEC had one of the first 64 bit CPUs on the market, and I spent a lot of my time fixing 64 bit issues in open source software (mostly assumptions that sizeof(int) == sizeof(void *)).
A few years later, I worked as a systems programmer for a department doing research using DEC Alphas, and wrote device drivers for OSF/1 (aka DEC UNIX) for Myrinet. I was lucky enough to get them to pay me to help Doug Rabson port FreeBSD to the DEC Alpha as part of my job. We ran FreeBSD/alpha on our research machines until I left.
I remember getting an API UP1000 direct from them so we could have FreeBSD board support for the irongate chipset from day one. I had that machine (running FreeBSD/alpha) as a personal desktop for several years. (this was before things like ACPI were a thing outside of x86, and each motherboard needed custom support)
I really, really miss SRM firmware. It was so much simpler than UEFI and openfirmware. I wish Intel had chosen it, rather than UEFI, to replace BIOS on x86.
EDIT: The 164LX is one of the few smaller PCI based Alphas that I never encountered in person. I think that was because DEC's discount structure made their workstations more affordable than the 164LX. Especially since we were running FreeBSD, and didn't need to pay for DEC UNIX.
I haven't really used SRM in 20 years, and I haven't used openfirmware in about the same amount of time. But if you boot me in front of an alpha running SRM and booting to the SRM>>> prompt, I could probably get it booted in a few minutes (show dev; boot dka3), but I've often flailed around on sparcs & ppc boxes running openfirmware, struggling to find disks and boot the os install media.
Then HP released (1991 or so) the "snake" workstations hp-730, about twice as fast as main sun competition, and also faster (at least on our codes) than the 21064/21164. Had a crazy fast X11 implementation. One of our collaborators cried that it was faster to run things from our workstation remote to his $30k sun workstation than it was to run it local.
HP rev'd the CPU in 1992 and brought out a HP-735, which was a cheap upgrade ($2,000 or so), which more than doubled the performance of the 730, added a much better SCSI system, doubled the clock, improved the cache, and improved the IPC. The floating point improvements were particularly noticeable.
I attended the "1000x" Alpha announcement in 1992. They promised a revolutionary 64 bit architecture promising 10x from IPC, 10x from increased cores, and 10x from increased clocks. Most marketing compared it to Sun's sparcs, which made it look good. Compiler optimization support looked uneven, sure some benchmarks looked good, but our computational workloads (EEG and MRI processing) preferred our HP-735.
Sadly around 1998 Dec (or maybe Compaq) decided not to sell the 264 through the same channels, so we ended up buying alphastation DS10s. But they were pricey, similar clock speeds, but around twice the performance of the 21164 because of various tweaks and moving to Out of Order execution.
Then not long after AMD brought out their answer to the Itanium, with the x86-64 extension. Suddenly PCs could do "real" work with a single process handling over 4GB of ram. Moving the memory controller on chip, so dual sockets has 2 memory systems and NUMA instead of Intel's shared front side bus was another revolution. Said FSB seemed largely responsible for the "Intel servers are slow" when busy. Disk I/O, network I/O, or using the second socket would substantially slow CPU performance.
All in all it was a pretty crazy decade that watched as the industry moved from single core, 100MHz, poor IPC, and 32 bit to multicore, 64 bit, over 1GHz, and much improved IPC. Sadly early 2000s were when Intel bluffed the industry into killing off alpha, pa-risc, (mostly) sparc, and mips for Itanium chips that shipped last, slow, and never hit their promised targets. For the first few years alpha matched itanium performance, despite having a huge process disadvantage.