| Episodes drawn on | 5 |
| Cited here | 5 |
| Claims extracted | 58 |
| Gathering | guest episode |
| Other themes | |
Retrocomputing and electronics archaeology concern the design, bring-up, failure modes and later restoration of computing and electronic hardware from the era of NMOS custom silicon, hand-taped printed circuit boards and dynamic RAM priced as the dominant line item in a bill of materials. The economics of early-1980s home computers were dominated by DRAM — a 49 dollar 16K machine was viable at a time when a tube of 64K DRAMs alone cost 99 dollars — which is why so much function was pushed into single custom chips.[222] Those machines remain instructive because their constraints were visible: memory contention between processor and video, marginal clock timing, metastability in custom silicon, and emissions governed by edge rate rather than clock rate all had to be solved with a handful of parts and no abstraction to hide behind.[222] Restoration and preservation add a second set of problems, from part scarcity and storage degradation to the chemistry of ageing insulation and the generation loss embedded in surviving media.[222][540][586]
Design under cost constraint
Commodore’s design culture under Jack Tramiel imposed a hard chip-count budget, expressed internally as a rule of no more than nine chips, which shaped how much function had to be pushed into custom silicon rather than glue logic.[222] Cost reduction at the bottom of the range went as far as a single-sided printed circuit board with wire jumpers, chosen because it saved about a dollar per unit and the jumpers could be fitted by the auto-insertion machinery already present in the line.[222] The TED line was positioned deliberately below the Commodore 64 rather than against it, targeting a 79 dollar price for the complete unit less television and drive; its video chip offered 121 distinct colours, short of a nominal 128 because eight shades of black are indistinguishable.[222] The Commodore 64 itself sold about 27 million units over roughly five years, a figure that stood as the record for a single computer model until tablet-class devices reached comparable volumes.[222]
Not every rule was worth keeping. A reset circuit built from a plain non-Schmitt inverter is unreliable on a slow supply ramp, and adding a tenth chip to fix it was judged worth breaking the nine-chip rule over.[222]
Constraints were sometimes commercial rather than technical. Sync polarity was used as a deliberate lock-in mechanism on 1980s home computers: contemporary third-party monitors expected either two positive-going syncs or two negative-going syncs, so specifying one of each forced buyers onto the manufacturer’s own monitor.[222]
Video and memory contention
The central problem of early video chips was memory contention. Any processor access to display memory while the video chip is fetching puts visible dots on the screen, so early designs confined processor access to the retrace intervals.[222] That left only about ten to thirteen microseconds during horizontal retrace, plus a longer window at vertical retrace, in which to touch display memory.[222]
The VIC chip avoided the restriction by exploiting the 6502’s fixed two-phase cycle: the bus is handed to the video chip during the first half of every cycle and taken back for the second, so the processor runs at full speed while the video chip gets nearly all the DRAM bandwidth it needs.[222] Terminology of the period was constrained by trademark rather than function — the term sprite was a Texas Instruments trademark at the time, so Commodore documentation had to call the same hardware feature a movable object block.[222]
Marginal timing in the Commodore 64 originated in an inadequate master clock frequency: there was no high enough frequency to divide down into the short RAS pulse and the row-to-column address switch that DRAM access requires, so the design barely worked in production.[222] Production floors under shipping pressure will patch such designs themselves, in this case adding capacitors on the RAS and CAS lines to make units pass, and may actively keep engineering from seeing it; metastability in a custom chip whose designer had not accounted for it was the underlying cause.[222]
Custom silicon
A semiconductor process recipe as written does not by itself give yield. Process engineers deliberately varied one parameter by a percent on each otherwise good run and plotted pairs of variables against each other, keeping any change that improved the result, so that the recipe drifted toward the yield optimum.[222] NMOS parts of the era were built in about seven layers including passivation, against the twenty or thirty layers of modern processes, which is part of why layer-level patching of a design was practical at all.[222]
Turnaround on custom chips was shortened by the 1-2-3 half-lot technique: half a wafer lot was run all the way through, and the other half stopped after the first three diffusion layers and put into storage, so a fix confined to the upper layers could be applied to the stored half and returned quickly instead of restarting a lot costing around 300,000 dollars.[222] Development schedules were correspondingly aggressive — the Commodore 128 was taken from concept to a CES demonstration in about five months with four or five custom chips being developed in parallel, each of which failed at least once during that period.[222]
Defects at the silicon level reached the field. A reverse-engineered PLA used in the Commodore 64 suffered a passivation defect visible under a microscope as a purple creeping corrosion under the passivation layer, and was a significant cause of field failures.[222] A back-bias generator added to a video chip made it worse rather than better: the on-chip voltage doubler drifted during intervals when the chip was not being accessed, producing a display whose leftmost character column was dark and which brightened progressively across the screen.[222] That revision could be salvaged at the package level, because the pin-one indicator slot in the middle of an old ceramic package is connected to the die substrate; soldering a wire into that slot and grounding it held the substrate bias steady and restored the previous revision’s behaviour.[222]
Thermal and marginal-timing problems were addressed similarly. Moving a hot custom chip’s heat into the shielding can was done by specifying a beryllium copper lead frame, which conducted the die’s heat out through the package into metalwork already present for emissions reasons.[222] A marginal custom chip can also be masked by cooling it, and a software author demonstrating on a chip cooled with an ice cube concealed the fact that the 80-column path did not actually work at temperature.[222] The eventual fix for that 80-column video chip was to phase-lock it to the main 14.318 MHz colour clock instead of running it from its own 16 MHz source, then shift the shift-register clock edge earlier or later, tuning each unit individually to the sweet spot where it worked.[222]
System bring-up
Adding a reset switch to a machine whose video chip performs DMA is not trivial, because resetting mid-DMA leaves the processor unable to start; the C128 solved it with back-to-back open-collector gates forming a latch that held the processor off until the DMA completed.[222] A peripheral that seized the bus at the reset vector prevented the machine starting at all, and the fix there was architectural: the Z80 was made to run first out of reset as a bootstrap that established the correct mode before handing control to the 6502.[222]
Bus sharing between processor families produced subtle electrical faults. Halting a Z80 sharing a 6502 bus at the wrong point in its clock and pause pattern leaves the output buffers enabled while the internal bus floats, so the pin drives an amplified intermediate level near a volt that cannot be pulled either way; one brand of address multiplexer read that as a low while another oscillated, which made the fault look like a parts speed problem.[222]
Integrating a function rather than shipping it on a cartridge changed the power budget by an order of magnitude: a CP/M cartridge drew about 0.6 A where the built-in version needed roughly 100 mA, which mattered because the C128 used a switching supply that shuts down when overloaded, whereas the C64’s potted linear supply merely sagged.[222]
The Commodore 128 also carried one of the first memory management units in a home computer, but it was specified without a supervisor mode, so the mapping registers were writable by user code and the system had no way to protect its own configuration.[222]
Backward compatibility
Backward compatibility can be broken by cosmetic changes. Cleaning up a character-generator ROM moved the dot on the letter i, and a paint program that scaled glyphs straight out of that ROM then filled the wrong region and painted the whole background.[222] The fix was to double the size of the character ROM, hold both the old and the corrected font, and select between them with a single address line driven high in 64 mode and low in 128 mode — initially by brick-laying a second ROM on top of the first.[222]
Removing a legacy register is riskier than hiding it, because a register that failed to reappear after a reset would leave the machine unrecoverable. The C128’s 2 MHz speed registers were therefore left present but placed where software was not supposed to write, and a game that incremented through 256 register values instead of decrementing through 24 duly wrote garbage across them.[222]
Both machines carried a full ROM monitor built in, so the machine itself provided an assembly-level development environment; internally the same feature was regarded as making the machine easy to copy software on.[222]
Emissions and interference
An aperture in a shield does more than let frequencies through the hole: a dipole forms in the metal immediately alongside the opening and radiates, which is why aperture placement and size, rather than the shield material, dominate emissions performance.[222] Emissions from a 1980s machine running a 16 MHz clock were still a problem at 180 MHz, because what radiates is set by the edge rate rather than the clock rate, and bus contention between devices fighting for the same bus adds its own noise signature.[222]
The same emissions are usable as a diagnostic. A cheap AM radio left running quietly beside a prototype makes the machine’s activity audible and characteristic enough that a failed ROM test can be recognised by ear before anything is on screen.[222]
Interference also travels out through connectors. Running processor data lines straight to a membrane keyboard and joystick ports exposes the bus at the connector: plugging in a joystick produced sparkles on the display and, held near the monitor, crashed the processor outright.[222]
Troubleshooting practice
Digital signals are best debugged as analogue events. A logic one is a waveform that attempted to reach a threshold and rang on the way, and treating it that way rather than as an abstract bit is what makes marginal timing and integrity faults visible.[222]
A large fraction of production board faults are visible rather than electrical: about half of a box of failed boards could be repaired on visual inspection alone, without ever powering them, by learning what a bad joint or misplaced part looks like.[222] Piece-rate consumer repair economics reward that pattern recognition over analysis — paid roughly ten dollars per set repaired, an hour of genuine troubleshooting loses money while recognising a previously seen fault and fixing it in ten minutes is profitable.[222]
Some faults only appear under deliberate stress. Holding a product’s mains input at the voltage where its reset circuit oscillates, by parking a Variac at that point, is a destructive stress test for battery-backed memory designs, and revealed that a whole population could be corrupted at that supply level.[222] Battery-backed static RAM in the pre-flash era was built from a cell battery and a diode on the supply rail, and the choice of which chip-enable line is used to deselect the part materially affects whether data survives the power transition; using the wrong one corrupts the contents.[222]
Precision analogue work of the period imposed its own discipline: digital scales built around a 6502 with a load-cell front end resolved one part in 50,000, a resolution at which ground loops must actually be understood and eliminated rather than worked around.[222]
Layout, tooling and regulation
The Commodore 128 was the company’s first CAD-produced printed circuit board, laid out on a Sci-Cards system driven from a VAX; the immediately preceding machines — the TED, Plus/4 and 116 — were still hand-taped, and boards designed in the United States were re-laid-out by Commodore Japan for auto-insertion in production.[222]
Late fixes were constrained by regulatory approval as much as by schedule. A change had to be made from leftover logic parts already on the board so it qualified as a class 1 permissive change and did not force a new FCC submission, since a machine shown at CES in January had to ship by May to reach retailers for Christmas.[222]
Entry into the field was gated by capital equipment. Before cheap tools and a secondhand market existed, microprocessor work required a development system costing ten to twenty thousand 1970s dollars plus an oscilloscope nobody could personally afford, which confined the work to people whose employer bought the equipment.[222]
Safety failures of the era could halt production outright: the TI-99 had a field failure in which transformer insulation broke down as the unit heated, energising the chassis to mains-derived potential.[222]
Displays of the period
Seven-segment LED displays of the 1970s and early 1980s were extremely dim for their power. Between 20 and 30 milliamps per segment produced barely visible output, and early Hewlett-Packard modules with the decoder bonded into the display ran hot enough to be alarming while delivering less light than a comparable filament lamp.[540] Incandescent displays consequently remained in production well into the 1980s and 1990s long after LEDs were available, for the specific reason that a filament display was the only technology readable in daylight.[540]
Preservation and restoration
Ageing materials give vintage equipment its characteristic smell, and the chemistry is identifiable: beeswax used to coat coils in radios, phenols outgassing from phenolic boards, and the plasticiser used in PVC insulation — the same compound responsible for the green goo that attacks copper and nickel as it ages.[540]
Stored components are not indefinitely safe either. Bare semiconductor parts have storage requirements, and chips that were not kept under nitrogen degraded in storage and were unusable when the remaining inventory changed hands.[222] Sourcing is further complicated by market behaviour: prices for obscure vintage display devices on the secondhand market are extremely sensitive to publicity, with tubes that changed hands for tens of dollars moving to hundreds or thousands once documented publicly, which is a real risk for anyone planning a restoration around a specific part.[586]
Keeping an obsolete analogue function alive does not force a choice between digital emulation and scavenged secondhand parts. The middle path is re-engineering the function from available parts at a higher component count, or committing to a custom ASIC or an FPGA that behaves like the original.[263] Analogue signal paths also persist in battery-powered audio equipment for concrete reasons rather than nostalgia: a microcontroller costs battery life and injects switching and clock noise into a circuit whose whole purpose is a low-noise, wide dynamic range path.[263]
Valve equipment imposes its own restoration procedure. Matching amplifier valves requires burn-in first, because transconductance does not settle until the tube has run for something like eight to twelve hours, so measuring a fresh tube gives a match that will not hold.[647] A production burn-in rig for valve matching kept every heater on continuously and cycled the racks under load in turn on a timer overnight, after which the tubes moved to a separate automated tester that ran each one through measurement cycles and sorted them into matched sets.[647]
Obsolete computing platforms are still commercially supported where installed equipment depends on them: industrial computer manufacturers continue to offer obsolete operating systems as a product line, including PC/104 form factor boards with DOS in ROM for instant boot.[647]
Media archives
Preservation limits also apply to recorded material. Content shot on film in the 1980s can be rescanned at 4K, while 1990s television shot on standard-definition videotape has no higher-resolution original to return to, so an entire decade of production is permanently limited to its original interlaced resolution.[540] Archive quality is usually limited by generation loss rather than by format: two-inch quad masters were routinely wiped and reused, leaving three-quarter-inch syndication copies of copies as the surviving source, and upscaling only works well when the source generation is genuinely clean.[540]
References
| Episode | Title | Date |
|---|---|---|
| 222 | An Interview With Bil Herd - Zany Z80 Zygology | October 27, 2014 |
| 263 | An Interview with Fran Blanche | August 19, 2015 |
| 540 | The Space Time Continuum with Fran Blanche | May 4, 2021 |
| 586 | Fran Blanche Version 3 | |
| 647 | Dave hanging with Fran Blanche | October 10, 2023 |