Author Topic: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe  (Read 1019 times)

0 Members and 1 Guest are viewing this topic.

Offline intabits

  • Regular Contributor
  • *
  • Posts: 237
  • Country: au
Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« on: October 19, 2020, 05:46:38 am »
This might normally belong in the vintage computing sub-forum, but I thought it would probably be of more general electronics interest.

My teardown of a Power Supply Assembly, one of several used in an IBM 3031 mainframe circa 1978.

Supply input is 3 Phase 208V at 415Hz.
This is the larger of the two units I have, the teardown is of the slightly smaller one which provides 2.3V at 250A.

Primary regulation is via 6 SCRs on the secondary windings, followed by what I presume is a transistor shunt regulator.




The IBM 3031 Central Processor




Overall schematic




I surmise that this is a 3phase bridge rectifier for internal power, and a final output shunt regulator.




My Teardown:-

 
The following users thanked this post: doktor pyta

Offline niconiconi

  • Regular Contributor
  • *
  • Posts: 61
  • Country: cn
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #1 on: October 19, 2020, 09:32:22 am »
A pure linear supply?!  :o I thought many minicomputers from the same era were already using a switching design (e.g. PDP-11), and mainframes can't be worse. Apparently I was mistaken.
« Last Edit: October 19, 2020, 09:38:02 am by niconiconi »
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 8038
  • Country: de
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #2 on: October 19, 2020, 09:36:12 am »
Interesting old style construction. I have never seen SCRs used with such a low voltage before.

Anyway it was new to me they used such a low voltage. TTL wants 5 V and ECL is also around 5 V or often more like -5 V.
Maybe the 3031 is still using core memory - this may need a relatively low voltage, but why some 250 A ?
 

Offline m k

  • Regular Contributor
  • *
  • Posts: 103
  • Country: fi
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #3 on: October 19, 2020, 09:48:37 am »
Back in the day it wasn't unheard of that following circuit board was short circuiting and on fire but power was just happily supplying for a bit higher load.
If memory serves VAX 11/750 or so had 1,8V or so power but it's output was only over 100A.

I was once about to try that kind of an emitter follower chain with NPNs only where control was also a stack.
With low input voltage and high output power it becomes quite easily quite nasty.
In my case last stage NPNs were 2N5302 with hfe 4 for 20A since 2N5686 was clearly too easy with that load and stuff of course from old bag of all sorts.
Never got far enough for heatsinks.
 

Offline David Hess

  • Super Contributor
  • ***
  • Posts: 12045
  • Country: us
  • DavidH
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #4 on: October 19, 2020, 04:30:45 pm »
A pure linear supply?!  :o I thought many minicomputers from the same era were already using a switching design (e.g. PDP-11), and mainframes can't be worse. Apparently I was mistaken.

The SCR preregulator is switching, but with 6 phases at 60 Hz.  Check out the size of that output inductor.

The linear shunt regulator's dissipation is not actually all that great compared to the rest and is required because the minimum load current is high.

I have seen bipolar transistor switching designs from that era in the same power range and I am not sure why they were not used here.  Maybe it was because of rectifier losses at such low output voltages?
 

Online coppice

  • Super Contributor
  • ***
  • Posts: 6023
  • Country: gb
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #5 on: October 19, 2020, 04:46:45 pm »
A pure linear supply?!  :o I thought many minicomputers from the same era were already using a switching design (e.g. PDP-11), and mainframes can't be worse. Apparently I was mistaken.
Many machines used switching supplies in 1978, but their reliability couldn't compete with a linear supply at that time. So, mainframes still used linear supplies. The larger machines used mains to 400Hz rotary converters in front of the linear supply. Those rotary converters could have a diesel motor added to them and some short term batteries, to provide uninterruptible power.
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 8038
  • Country: de
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #6 on: October 19, 2020, 08:18:27 pm »
With the SCR part the regulator is not fully linear. One can see this from the heat sink. I would estimate a maximum power of some 100-200 W for the transistors. This would correspond to only some 0.4 -1 V of effective drop a series regulator would have to work with. Even the shunt regulator part works in combination with the choke, so it only has to care about the ripple current and fast load transients.

SCRs are unusual for such a low voltage, but they are available for high currents.

It is a neat way to get around having large electrolytic caps. Inductors as energy storage can be quite effective at higher power. Like a transformer they scales slightly faster than linear in weight and they also scale well to low voltages.
 
The following users thanked this post: niconiconi

Offline intabits

  • Regular Contributor
  • *
  • Posts: 237
  • Country: au
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #7 on: October 20, 2020, 06:25:11 am »
The SCR preregulator is switching, but with 6 phases at 60 Hz.  Check out the size of that output inductor.
Maybe it was because of rectifier losses at such low output voltages?

It's actually 415Hz, so the ripple is 2490Hz.
My limited testing of the inductor seems to indicate it is only about 60-100uH. It has only about 15 turns of wide flat copper.

The rectifier loss is a good point that hadn't occurred to me. I wonder how the voltage drop across an SCR compares with a diode...
 

Offline intabits

  • Regular Contributor
  • *
  • Posts: 237
  • Country: au
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #8 on: October 20, 2020, 06:36:03 am »
Anyway it was new to me they used such a low voltage. TTL wants 5 V and ECL is also around 5 V or often more like -5 V.
Maybe the 3031 is still using core memory - this may need a relatively low voltage, but why some 250 A ?

There were several (at least 4) of these supplies in the 3031, of various voltages, but I don't know what voltages  the others were.

IBM stopped using core memory in about 1970 with System/370, which this was a successor to.
The 3031 definitely used semiconductor memory, though I'm not sure if it was static or dynamic.

There was a lot of logic in those frames, so not too hard to see where all those amps went. 
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 8038
  • Country: de
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #9 on: October 20, 2020, 08:31:07 am »
The SCR preregulator is switching, but with 6 phases at 60 Hz.  Check out the size of that output inductor.
Maybe it was because of rectifier losses at such low output voltages?
My limited testing of the inductor seems to indicate it is only about 60-100uH. It has only about 15 turns of wide flat copper.

The rectifier loss is a good point that hadn't occurred to me. I wonder how the voltage drop across an SCR compares with a diode...


At the high current a 100 ┬ÁH inductor can still store quite some energy (some 3 J at 250 A) - about comparable to a 1 F capacitor at 2.3 V. In actual use the change in current  / voltage would be smaller, still about comparable.
With the shunt regulator sinking some 40 A peak and some 4 ms for the ripple period this would be 10 A/ms and thus 1 V drop (working range) at the inductor. So the values looks reasonable.

I don't know the details on the motor / generator part. If wanted one can already do some coarse voltage stabilization with the generator and use it as a kind of magnetic amplifier.

The SCRs usually drop some 1.5 to 2 V depending on the current. It may be a little better with special low voltage types, but AFAIK there are 2 diode drops included, so hard to get below some 1.2 V.  So I would consider the SCR loss about twice the loss of a diode, maybe a little higher. Still the SCR replaces a diode and a switching device and they are available for the high currents. At high currents a transistor also has a CE saturation voltage in the 0.5 to 1 V range.
 
The following users thanked this post: intabits

Online coppice

  • Super Contributor
  • ***
  • Posts: 6023
  • Country: gb
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #10 on: October 20, 2020, 02:17:47 pm »
Anyway it was new to me they used such a low voltage. TTL wants 5 V and ECL is also around 5 V or often more like -5 V.
Maybe the 3031 is still using core memory - this may need a relatively low voltage, but why some 250 A ?

There were several (at least 4) of these supplies in the 3031, of various voltages, but I don't know what voltages  the others were.

IBM stopped using core memory in about 1970 with System/370, which this was a successor to.
The 3031 definitely used semiconductor memory, though I'm not sure if it was static or dynamic.

There was a lot of logic in those frames, so not too hard to see where all those amps went.
The first 370s still used core memory. It was replaced by semiconductor memory a little later. IBM were a real powerhouse of semiconductor technology then, but only made silicon for their own use. So, its not always easy to know the exact technology they were using. This also explains the funky voltages of the power supplies in this thread. They had little need to follow industry norms. The first 370 with semiconductor memory shipped in 1971. The first commercial DRAM chip (Intel 1103) appeared a few months earlier, but in the early days many people had reliability troubles with DRAM, so its not clear that IBM would have used DRAM in a system where reliability was of great importance.
 
The following users thanked this post: intabits

Offline Miyuki

  • Frequent Contributor
  • **
  • Posts: 434
  • Country: cz
    • Me on youtube
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #11 on: October 20, 2020, 02:40:12 pm »
A pure linear supply?!  :o I thought many minicomputers from the same era were already using a switching design (e.g. PDP-11), and mainframes can't be worse. Apparently I was mistaken.

The SCR preregulator is switching, but with 6 phases at 60 Hz.  Check out the size of that output inductor.

The linear shunt regulator's dissipation is not actually all that great compared to the rest and is required because the minimum load current is high.

I have seen bipolar transistor switching designs from that era in the same power range and I am not sure why they were not used here.  Maybe it was because of rectifier losses at such low output voltages?

Maybe they want to avoid low frequency switching supply, also noise and reliability can be a big issue

I have one old "portable" oscilloscope so it works from 12V supply and has internal SMPS but it works on 5-6kHz range
It is a terrible experience to turn it on, it can be very annoying sound

btw Father used to work with mainframes back then and have some funny/terrible stories, as when in case of short circuit supplies like those can happily burn cables all the way to that short and must be shut down manually

But compare it with today computers, any high-end GPU can easy have buck regulator with a higher current than this, it is insane when thinking about it
And feeding it into single chip
 

Offline m k

  • Regular Contributor
  • *
  • Posts: 103
  • Country: fi
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #12 on: October 21, 2020, 03:40:41 pm »
But compare it with today computers,

I was once trying to use a machine with double 20" belt driven Winchester where one had a bearing problem.
Doable with earmuffs but after hours only, it wasn't a residential area and the drive still operated normally.
 

Offline jmelson

  • Super Contributor
  • ***
  • Posts: 1663
  • Country: us
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #13 on: October 21, 2020, 11:29:18 pm »
Interesting old style construction. I have never seen SCRs used with such a low voltage before.

Anyway it was new to me they used such a low voltage. TTL wants 5 V and ECL is also around 5 V or often more like -5 V.
Maybe the 3031 is still using core memory - this may need a relatively low voltage, but why some 250 A ?
ECL is a power hungry type of logic.  IBM used what they called MST IV (Monolithic Solid Technology) quite similar to other commercial ECL families.
But, they did things a little differently.  They used supplies of -3 V and +1.2 V, at least on the earlier versions.  This allowed the logic swing to be balanced around ground, so you could pull a terminator off a bus and bring it out to a scope for examination.  Because the ECL chips could only pull up toward the positive supply (the outputs were the emitters of NPN transistors) you had to provide pull-down with a hefty resistor.  To assure good logic levels for the logic-low state, the pull-down resistor had to be a lower resistance than would be the case when the termination was to a negative supply.

The IBM 370/145 (a sizteen-bit hardware implementation of a 370 architecture) had two supplies, + 1.2 V and -3 V, both rated at 390 A!  So, 1600 W for a 16-bit mini.  (That includes power for a lot of fast static ram control store.)

The 3031 was a deriviative of the IBM 370/168 and used pretty much the same logic technology, as far as I know.

Jon
 

Offline jmelson

  • Super Contributor
  • ***
  • Posts: 1663
  • Country: us
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #14 on: October 21, 2020, 11:37:52 pm »

The first 370s still used core memory. It was replaced by semiconductor memory a little later. IBM were a real powerhouse of semiconductor technology then, but only made silicon for their own use. So, its not always easy to know the exact technology they were using. This also explains the funky voltages of the power supplies in this thread. They had little need to follow industry norms. The first 370 with semiconductor memory shipped in 1971. The first commercial DRAM chip (Intel 1103) appeared a few months earlier, but in the early days many people had reliability troubles with DRAM, so its not clear that IBM would have used DRAM in a system where reliability was of great importance.
The IBM 360/85 was the first mainframe they delivered with monolithic ICs.  It had a cache (storage buffer in IBM lingo) made of 16-bit ECL static RAM, and also had a 500-word writeable control store in the same technology.  The 370/165 was basically the same machine, but those memories were redone in 64-bit chip technology, and all control store was writeable.  The 360/85 was a rare model, the only ones delivered outside IBM were to the NSA.
The 370/165 and 370/168 were quite popular machines.  The first two 370 models (/165 and /155) used core memory.  The versions of those machines with vircual memory (Dynamic Address Translation in IBM parlance) wre the /168 and /158.

The 370/145 was the first machine with solid-state main storage, they did use static RAM for these, but it was a MOS technology, not ECL as was used for the faster memories for caches and control store.  The problem with DRAM is how to schedule the refresh cycles when so many devices are accessing the memory.  Of course, later, people figured out how to do it just fine.

Jon
 

Online coppice

  • Super Contributor
  • ***
  • Posts: 6023
  • Country: gb
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #15 on: October 21, 2020, 11:50:36 pm »
The 370/145 was the first machine with solid-state main storage, they did use static RAM for these, but it was a MOS technology, not ECL as was used for the faster memories for caches and control store.  The problem with DRAM is how to schedule the refresh cycles when so many devices are accessing the memory.  Of course, later, people figured out how to do it just fine.
Refresh was never a real problem with DRAM. It was just something you needed to do. The key problem with early DRAM for high reliability systems was a significant soft error rate. Even with error correcting memory there was a significant risk of errors. This exploded when we got to 64k devices, at which point it had to be addressed if the DRAM market was to move forwards. It was, and we don't seem to have had a significant soft error issue since that time.
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 8038
  • Country: de
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #16 on: October 22, 2020, 06:52:34 pm »
Something like +1.2 and -3 V could make sense for ECL, it is not too far from the normal 5 V. I would imagine one would still have a single 4.2 V supply for most of the power and get only the difference from a separate supply or virtual ground.

AFAIR the soft errors where still an issue with the PCs at around the 386/486. There was the division into just parity for more private use and error corrected memory (36 instead of 32 bits) for workstations and servers. Later and still now the DRAM chips include error correction already internally. This step gave the DRAM quite some increase in reliability.
 

Offline jmelson

  • Super Contributor
  • ***
  • Posts: 1663
  • Country: us
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #17 on: October 23, 2020, 06:27:01 pm »
Refresh was never a real problem with DRAM.
Well, if you speced out the system so several channels transferring at once could use up 100% of the main storage bandwidth, then there was no time available for the refresh, and it would cause machine checks when the memory was not available in time.  IBM mainframes lived and died by the memory bandwidth, and so they didn't want to have to limit the system to leave time for refresh.

Jon
 

Offline jmelson

  • Super Contributor
  • ***
  • Posts: 1663
  • Country: us
Re: Teardown of 2.3V 250A PSU from a 1978 IBM mainframe
« Reply #18 on: October 23, 2020, 06:30:35 pm »
Something like +1.2 and -3 V could make sense for ECL, it is not too far from the normal 5 V. I would imagine one would still have a single 4.2 V supply for most of the power and get only the difference from a separate supply or virtual ground.
Many systems did this, Motorola recommended running their ECL with Vcc grounded and Vee at -5.2 for MECL 10K and -4.5 for MECL 100K families.
That apparently gave the best noise margins.  Then, termination was to -2.0 V from a separate power supply.  But, that was only one way to set up an ECL system.

Jon
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf