Bob Supnik is a really interesting guy. Probably best known these days for being a maintainer of the SIMH emulation environment. He's forgotten more about computers than most people will ever know.
The simh maintainer kicked Bob out a couple years back in a spat over containerization (no, not that kind) and changed the simh license to non-Free over it. A new fork has emerged that Bob is involved in, but it's not the same simh that can trace its lineage back to DEC SCP.
The trouble with the CISC vs RISC interpretation is that the x86 was CISC but trashed all other architectures in the 1990s especially RISC architectures including
I think it’s as simple as a mass market architecture that sells more units can justify more investment than anything that sells fewer units. If DEC had been able to steal market share from the 386 with a VAX based product it might have been able to take the 386’s place but from a business perspective they didn’t want to cannabalize sales of higher margin minicomputers. The transition from bipolar to CMOS was also difficult because it did mean a regression in performance, IBM addressed this in the 390 by introducing a clustering solution but it was a bold and risky move.
Alpha died because HP decided that Itanium was the one true processor and killed PA-RISC and Alpha (which it had acquired) in favor of itanic. I'd love to live in a world where Alpha had been free to compete on merit. Who knows, maybe we'd all be running Alpha desktops and laptops.
DEC was (slowly) on its way to making alpha mass-market. Around the time of the Compaq acquisition, they started offering the CPU to 3rd parties to design their own boards. One example is the API UP1000, which had an AMD irongate chipset and an EV6 CPU. (I had an early sample that I ported FreeBSD/alpha to). Alpha was the only somewhat popular non-x86 platform that Windows NT ran on. FX!32 made it possible to run x86 apps, like Rosetta on Apple M4. Alpha was the first non-x86 linux port.. I was at the USENIX where John "Maddog" Hall gave Linus an Alpha (I think it was a Multia) Alpha was also the first 64-bit FreeBSD port (which I contributed to), and paved the way for amd64.
Alpha had a HUGE amount of momentum when HP summarily executed it. I still hate HP with a passion, and I still won't buy anything from them, 20+ years later.
I have heard that the Alpha architecture maintainers said that their options for improvements were limited, and themselves advocated ending development.
What I also know is that DEC chose ARM for low-power applications, because the design of the Alpha was simply not capable of scaling to lower power usage.
"According to Allen Baum, the StrongARM traces its history to attempts to make a low-power version of the DEC Alpha, which DEC's engineers quickly concluded was not possible."
I suspect it was less "Alpha is not capable of being low power" and more "We didn't have resources to make low-power from-scratch implementation" including the transistor budget driven by available lithography and constrained by Alpha being pure 64bit architecture.
Most importantly, outside of packaging the core in SoC-like setup, Digital essentially could only ship one microarchitecture at the time with Alpha.
> I have heard that the Alpha architecture maintainers said that their options for improvements were limited, and themselves advocated ending development.
Interesting. One of the main design principles for the Alpha was for longevity, and to have a thousand-times increase in performance over 25 years.[0]
I'm only halfway through the Mashey thread, but this is an interesting assertion, leading to what I think can and cannot be done (in my profound state of ignorance):
"ONE LAST TIME: it wasn't politics that ended the VAX, it was
engineering judgement by excellent (IMHO) DEC engineers, and the
economics of doing relatively low-volume chips [low-volume compared to
X86]. Nobody could afford to keep the VAX ISA competitive at DEC
volumes...."
In contrasting this against Intel, what x86 could always obtain was and is performance. A large base of buyers will enable architectural improvements.
As an example, Bob Colwell was brought in to implement out of order opcode/microcode execution in the Pentium Pro. This was contemplated for the VAX, but never seriously entertained.
What could not be added to the Alpha was power efficiency, and x86 finds itself regularly beaten in this race as well.
It seems that money can buy only some things. Performance can be bought, but efficiency cannot.
___
EDIT: I have read this before, but it's always funny.
MIPS was the most provocative and frightening thing in high performance computing in the 80's and early 90's.
"IBM, DEC, and HP were among the few that actually had the expertise to
do CMOS micro technology, albeit not without internal wars. [I was
invovled in dozens of these wars. One of the most amusing was when IBM
friends asked me to come and participate in a mostly-IBM-internal
conference at T. J. Watson, ~1992. What they really wanted, it
turned out, was for somebody outside IBM politics (i.e., "safe") to
wave a red flag in front of the ECL mainframe folks, by showing a
working 100Mhz 64-bit R4000.]"
I was struck by that quote, too. Set aside IBM inventing RISC in the first place. How was it possible that in 1992, two years after RS/6000 hit the market and made IBM a first-class RISC workstation player—a peer of Sun and HP in terms of tech and sales—did IBM people still feel the need to resort to the subterfuge of a non-IBM engineer showing off a non-IBM RISC CPU to the IBM mainframe folks to give them a kick in the butt? The mind boggles.
x86 isn’t a VAX though, not all CISC architectures are equally complex (or RISC arch’s reduced), and VAX does have a reputation for being a particularly CISCy CISC. We don’t really know fully if the same tricks would have worked as well with it.
And, in fact, Mashey specifically discusses what you identified, and what Houle wrote (rushing to post a gotcha on HN, obviously without having read Mashey's lengthy writings).
Yeah, always good to read it. He goes into some detail about a hypothetical VAX-based x86 killer, tho some of it is "when I had beers with the guys designing the VAX..." reminiscing.
Comp.arch was something really special. Guys like Mash, John McCalpin, Bob Colwell, Eugene Maya, Mitch Alsup, Terje Mathisen...folks who really, really understood computers, what the real tradeoffs are, did actual homework instead of guessing, and were generally able to discuss the issues without being jerks. Good times.
What I've read was that DEC had a huge amount of regret over the PDP-11 having too small of an address space. It could be that experience led them to think the answer to their problems was to be early to market in the 64-bit age with the Alpha. They did have VMS for the Alpha and later Win NT but high-powered RISC processors were a crowded space in the 1990s.
If I remember correctly, it was basically 64bit because Alpha was basically a (good) virus inside of DEC. Or maybe like secret society that revolted against the leader of the company.
Olson had killed 32 bit PRISM and they already had VAX that was 32 bit and people making processor for it. To get people all over DEC to buy into Alpha (Alpha barley had any budget of it own) it had to be something new, and winning 64 bit did make sense.
Olson really killed basically everything that make sense, that DEC survived so long with Olson as CEO is kind of crazy. The amount of horrible decision starting as early as the early 70s is kind of crazy.
VAX had also been early, driven by Gordon Bell not Olson, very few of the competitors had 32 bit processors then, and people like Data General and Prime struggled to develop them in response. Funny enough a hardware guy on the VAX team basically proposed RISC-like architecture but it was rejected because they optimized for code size. To bad that they didn't hit on the idea of compressed instructions.
x86 is closer to RISC designs than "respectable CISC" of the era that defined the split between RISC and CISC. It's why huge chunk of instructions on modern x86 can execute in essentially one cycle - the encoding is not all that complex, the instructions are also not all that complex, and majority of "complex addressing" can be pipelined into having an extra stage capable of doing "LEA" instruction on it's own, unlike either VAX or 68k
At the point Alpha got killed, it was the fastest option, and then HP had to resurrect it anyway because the supposed Alpha killer (Itanium) was slower in practice. That said, the killing was done by Compaq, not HP - and done fully on politics/vibes not actual thorough thinking. A lot of late Alpha design ended up being "inherited" in ways by later x86 cores - Athlon reused chunks of EV6 architecture which also enabled cheap Alpha motherboards (but made Athlon-MP extra expensive, unfortunately). K8 heavily drank from experience of the team members who worked on EV7 which exhibited the same fabric structure.
I don't think that x86 beat Alpha. Alpha was more expensive because of a bunch of reason, but just in terms of how much performance they got out of the same amount of transistors, Alpha was better.
But I think generally speaking you are right, the size of the teams that all the RISC people had compared to the amount of resources Intel had was not really comparable.
> The transition from bipolar to CMOS was also difficult because it did mean a regression in performance
DEC already had the VAXCluster by 1983 and they were working on a bunch of implementations at the same time. They did a lower performance VAX already when the did a AMD 29000 bit-slice implementation in 1982. So they did know the value lower performance. By 1985 they had matched the 780 in a single chip. Basically by 1988 the CMOS version was already faster then anything else.
The transition was really just difficult because some people (who happen to be CEO) didn't want to see the reality.
If you want to learn about the history of VAX chips I very strongly recommend this oral history:
Supnik, Robert oral history
https://www.computerhistory.org/collections/catalog/10273826...
If you have any interest in VAX or DEC or chip design in the 80s. This is a must watch.
Later also goes into how Alpha was created in Part 2.
Bob Supnik is a really interesting guy. Probably best known these days for being a maintainer of the SIMH emulation environment. He's forgotten more about computers than most people will ever know.
The simh maintainer kicked Bob out a couple years back in a spat over containerization (no, not that kind) and changed the simh license to non-Free over it. A new fork has emerged that Bob is involved in, but it's not the same simh that can trace its lineage back to DEC SCP.
That's why I love Computer History Oral interviews the knowledge is crazy.
Another link for some history from the DEC side is this [1].
[1] https://simh.trailing-edge.com/dsarchive.html
The trouble with the CISC vs RISC interpretation is that the x86 was CISC but trashed all other architectures in the 1990s especially RISC architectures including
https://en.wikipedia.org/wiki/DEC_Alpha
and Intel’s own
https://en.wikipedia.org/wiki/Itanium
I think it’s as simple as a mass market architecture that sells more units can justify more investment than anything that sells fewer units. If DEC had been able to steal market share from the 386 with a VAX based product it might have been able to take the 386’s place but from a business perspective they didn’t want to cannabalize sales of higher margin minicomputers. The transition from bipolar to CMOS was also difficult because it did mean a regression in performance, IBM addressed this in the 390 by introducing a clustering solution but it was a bold and risky move.
Alpha died because HP decided that Itanium was the one true processor and killed PA-RISC and Alpha (which it had acquired) in favor of itanic. I'd love to live in a world where Alpha had been free to compete on merit. Who knows, maybe we'd all be running Alpha desktops and laptops.
DEC was (slowly) on its way to making alpha mass-market. Around the time of the Compaq acquisition, they started offering the CPU to 3rd parties to design their own boards. One example is the API UP1000, which had an AMD irongate chipset and an EV6 CPU. (I had an early sample that I ported FreeBSD/alpha to). Alpha was the only somewhat popular non-x86 platform that Windows NT ran on. FX!32 made it possible to run x86 apps, like Rosetta on Apple M4. Alpha was the first non-x86 linux port.. I was at the USENIX where John "Maddog" Hall gave Linus an Alpha (I think it was a Multia) Alpha was also the first 64-bit FreeBSD port (which I contributed to), and paved the way for amd64.
Alpha had a HUGE amount of momentum when HP summarily executed it. I still hate HP with a passion, and I still won't buy anything from them, 20+ years later.
I have heard that the Alpha architecture maintainers said that their options for improvements were limited, and themselves advocated ending development.
What I also know is that DEC chose ARM for low-power applications, because the design of the Alpha was simply not capable of scaling to lower power usage.
"According to Allen Baum, the StrongARM traces its history to attempts to make a low-power version of the DEC Alpha, which DEC's engineers quickly concluded was not possible."
https://en.m.wikipedia.org/wiki/StrongARM
Power efficiency was soon to become a chief concern. This alone would have ended the Alpha.
I suspect it was less "Alpha is not capable of being low power" and more "We didn't have resources to make low-power from-scratch implementation" including the transistor budget driven by available lithography and constrained by Alpha being pure 64bit architecture.
Most importantly, outside of packaging the core in SoC-like setup, Digital essentially could only ship one microarchitecture at the time with Alpha.
> I have heard that the Alpha architecture maintainers said that their options for improvements were limited, and themselves advocated ending development.
Interesting. One of the main design principles for the Alpha was for longevity, and to have a thousand-times increase in performance over 25 years.[0]
[0] https://danluu.com/dick-sites-alpha-axp-architecture.pdf
I'm only halfway through the Mashey thread, but this is an interesting assertion, leading to what I think can and cannot be done (in my profound state of ignorance):
"ONE LAST TIME: it wasn't politics that ended the VAX, it was engineering judgement by excellent (IMHO) DEC engineers, and the economics of doing relatively low-volume chips [low-volume compared to X86]. Nobody could afford to keep the VAX ISA competitive at DEC volumes...."
In contrasting this against Intel, what x86 could always obtain was and is performance. A large base of buyers will enable architectural improvements.
As an example, Bob Colwell was brought in to implement out of order opcode/microcode execution in the Pentium Pro. This was contemplated for the VAX, but never seriously entertained.
What could not be added to the Alpha was power efficiency, and x86 finds itself regularly beaten in this race as well.
It seems that money can buy only some things. Performance can be bought, but efficiency cannot.
___
EDIT: I have read this before, but it's always funny.
MIPS was the most provocative and frightening thing in high performance computing in the 80's and early 90's.
"IBM, DEC, and HP were among the few that actually had the expertise to do CMOS micro technology, albeit not without internal wars. [I was invovled in dozens of these wars. One of the most amusing was when IBM friends asked me to come and participate in a mostly-IBM-internal conference at T. J. Watson, ~1992. What they really wanted, it turned out, was for somebody outside IBM politics (i.e., "safe") to wave a red flag in front of the ECL mainframe folks, by showing a working 100Mhz 64-bit R4000.]"
I was struck by that quote, too. Set aside IBM inventing RISC in the first place. How was it possible that in 1992, two years after RS/6000 hit the market and made IBM a first-class RISC workstation player—a peer of Sun and HP in terms of tech and sales—did IBM people still feel the need to resort to the subterfuge of a non-IBM engineer showing off a non-IBM RISC CPU to the IBM mainframe folks to give them a kick in the butt? The mind boggles.
x86 isn’t a VAX though, not all CISC architectures are equally complex (or RISC arch’s reduced), and VAX does have a reputation for being a particularly CISCy CISC. We don’t really know fully if the same tricks would have worked as well with it.
And, in fact, Mashey specifically discusses what you identified, and what Houle wrote (rushing to post a gotcha on HN, obviously without having read Mashey's lengthy writings).
Yeah, always good to read it. He goes into some detail about a hypothetical VAX-based x86 killer, tho some of it is "when I had beers with the guys designing the VAX..." reminiscing.
Comp.arch was something really special. Guys like Mash, John McCalpin, Bob Colwell, Eugene Maya, Mitch Alsup, Terje Mathisen...folks who really, really understood computers, what the real tradeoffs are, did actual homework instead of guessing, and were generally able to discuss the issues without being jerks. Good times.
What I've read was that DEC had a huge amount of regret over the PDP-11 having too small of an address space. It could be that experience led them to think the answer to their problems was to be early to market in the 64-bit age with the Alpha. They did have VMS for the Alpha and later Win NT but high-powered RISC processors were a crowded space in the 1990s.
You can learn all about the details of how and why Alpha was developed:
https://www.computerhistory.org/collections/catalog/10273826...
Its in Part 2.
If I remember correctly, it was basically 64bit because Alpha was basically a (good) virus inside of DEC. Or maybe like secret society that revolted against the leader of the company.
Olson had killed 32 bit PRISM and they already had VAX that was 32 bit and people making processor for it. To get people all over DEC to buy into Alpha (Alpha barley had any budget of it own) it had to be something new, and winning 64 bit did make sense.
Olson really killed basically everything that make sense, that DEC survived so long with Olson as CEO is kind of crazy. The amount of horrible decision starting as early as the early 70s is kind of crazy.
VAX had also been early, driven by Gordon Bell not Olson, very few of the competitors had 32 bit processors then, and people like Data General and Prime struggled to develop them in response. Funny enough a hardware guy on the VAX team basically proposed RISC-like architecture but it was rejected because they optimized for code size. To bad that they didn't hit on the idea of compressed instructions.
x86 is closer to RISC designs than "respectable CISC" of the era that defined the split between RISC and CISC. It's why huge chunk of instructions on modern x86 can execute in essentially one cycle - the encoding is not all that complex, the instructions are also not all that complex, and majority of "complex addressing" can be pipelined into having an extra stage capable of doing "LEA" instruction on it's own, unlike either VAX or 68k
At the point Alpha got killed, it was the fastest option, and then HP had to resurrect it anyway because the supposed Alpha killer (Itanium) was slower in practice. That said, the killing was done by Compaq, not HP - and done fully on politics/vibes not actual thorough thinking. A lot of late Alpha design ended up being "inherited" in ways by later x86 cores - Athlon reused chunks of EV6 architecture which also enabled cheap Alpha motherboards (but made Athlon-MP extra expensive, unfortunately). K8 heavily drank from experience of the team members who worked on EV7 which exhibited the same fabric structure.
I don't think that x86 beat Alpha. Alpha was more expensive because of a bunch of reason, but just in terms of how much performance they got out of the same amount of transistors, Alpha was better.
But I think generally speaking you are right, the size of the teams that all the RISC people had compared to the amount of resources Intel had was not really comparable.
> The transition from bipolar to CMOS was also difficult because it did mean a regression in performance
DEC already had the VAXCluster by 1983 and they were working on a bunch of implementations at the same time. They did a lower performance VAX already when the did a AMD 29000 bit-slice implementation in 1982. So they did know the value lower performance. By 1985 they had matched the 780 in a single chip. Basically by 1988 the CMOS version was already faster then anything else.
The transition was really just difficult because some people (who happen to be CEO) didn't want to see the reality.
The 730 used the AMD 2900 (not 29000) chipset.
You are right, thank you for the correction.