Skip to content

Design review, datapath cards (control-path pass) — 2026-09-21

One card of the 2026-09-21 datapath review: the memory card v1.3, findings M1-M8. The other cards, the system-level items S1-S3 and the summary: hardware/reviews/datapath-2026-09-21.md.

Second pass after mechanical-2026-09-21.md (mechanical). Cards: ALU v3.2, Index Register 1.1, Memory v1.3, Mem Switch 1.1, Mem Register 1.0. Method: every select / enable / strobe / clock / reset net was traced pin-to-pin from kicad/<ver>/reports/netlist.net (pin names and types from the Eagle symbols) with chip types from the Eagle .brd, and checked against the way the sequencer microcode actually drives the bus (firmware/microcode/ucode-generator2/), because several hardware choices only work under a microcode convention. Signal polarity: - prefix = active low. Gate delays quoted are LS-TTL typicals from the datasheets in docs/datasheets/.

Severity: HIGH = would stop the card working or damage parts; MED = unreliable / depends on a convention that is not enforced by hardware; LOW = housekeeping or documentation.


Memory card v1.3 (traced on hardware/cards/memory/kicad/deprecated/v1.3-do-not-use/, then filed as kicad/v1.3)

Note 2026-09-24: this review traced what was then kicad/v1.3 (now kicad/deprecated/v1.3-do-not-use/), which is an earlier save of the design, not the card that was fabricated (hardware/cards/memory/eagle/v1.3/, KiCad kicad/v1.3/; filed first as v1.3-fusion-export-2026-09-24). The one circuit difference: on the built card IC5's enable (pin 19) is IC15 pin 12 = AND(-LO-RAM, -HI-RAM, -ROM-CS), not -VMA, so where the findings below say "IC5 enable" read "the chip selects, through IC15". The findings themselves (M1-M8) are unchanged by that: every pin they name is wired the same on the built card.

M1 — HIGH (masked by a microcode hack): FORCE-ROM clears on any -VMA cycle that starts with the address bus floating

Trace of the FORCE-ROM flip-flop clock:

  • IC12A (74LS74): PRE(4) = -RESET (X1.C30), CLR(1) = VCC, D(2) = GND, Q(5) = FORCE-ROM, CLK(3) = N$19.
  • N$19 = IC10.8 (74ALS00) = NOT(N$20); N$20 = IC10.6 = NAND(ADDR15 [X1.A18, raw bus line into IC10.4], N$22); N$22 = IC10.11 = NOT(N$25); N$25 = IC3.3 (74LS32) = -VMA OR -BUS-EN. So CLK = ADDR15 AND VMA AND BUS-EN, three gate delays (~35 ns) after -VMA falls.
  • The address is only driven onto the bus while -VMA is low: on the register card IC40.3 (74LS32) = -BUS-EN OR -VMA enables IC38 (74LS139, second half, pin 15) -> J3 -> IC38 first half (pin 1) -> -Rx-ADDRSEL -> IC41/42 (74LS244) output enables. That is ~75 ns typical (LS32 + two LS139 stages + 244 enable) before ADDR15 is driven.
  • Between memory cycles nothing drives ADDR0-15 (no pull-ups or pull-downs on the backplane: bus/backplane v2.0 contains only connectors and capacitors) and the memory card's LS inputs on ADDR15 (IC10.4, IC9.17) float high.

Failure scenario: on the first -VMA after reset, ADDR15 reads 1 (floating) for the ~40 ns before the register card drives it low, so N$19 produces a ~40 ns positive pulse (LS74 needs 25 ns) and clocks D=0 into IC12A: FORCE-ROM is cleared on the very first fetch regardless of the address, and the CPU fetches from uninitialised low RAM instead of the ROM. The bus-tester check of 2026-09-18 passed because the tester drives the address statically before it lowers -VMA.

The designer met this: ucode-generator2/main.c:102 and :115 (setSignal("-VMA"); // Hack prevent ROM mapping from triggering) assert -VMA in every microcode line with ADDR-REG-ID defaulting to PC, so the address bus never floats while the machine runs. The hack works, but it removes the -VMA qualification from every chip select on this card (the whole point of the v1.2 changes: IC7 G2A, IC10 -LO-RAM, IC5 enable are now permanently active), keeps IC5 (74LS245) driving BDATA from the bus at all times, and the hazard returns the moment any microcode line drops -VMA or the bus tester drives -VMA with the address lines tri-stated.

Bench: scope IC12 pin 3 against pin 5 on the first cycle after reset with the register card fitted and a microcode image that does not hold -VMA; or with the bus tester leave ADDR0-15 as inputs, hold -BUS-EN low, pulse -VMA low, and read FORCE-ROM (IC12.5): it drops.

M2 — MED: the 28C64 EEPROM is write-enabled by every -MEM-WR, including during FORCE-ROM

  • IC13 (28C64) -WE(27) = N$4 = IC6.4 = NOT(IC6.8 = NOT(-MEM-WR)): -MEM-WR re-buffered, no gating, no write-protect jumper. -CE(20) = -ROM-CS = IC6.12 = NOT(IC18.8), IC18 (74ALS30) fed by the ROM row of the U$1 block jumpers (pull-ups RN7).
  • While FORCE-ROM = 1, IC11 (74LS157, select pin 1 = FORCE-ROM, B inputs = VCC) forces BADDR12-15 = 1111, so every address selects block $F (IC7 Y7 -> jumper -> IC18 -> -ROM-CS).

Failure scenario: a write to any address while FORCE-ROM is still set (e.g. boot code initialising a RAM variable before its first jump above $8000), or any later stray write into $E000-$FFFF, writes the EEPROM. A 28C64 then spends ~10 ms in an internal write cycle during which reads return the toggle/poll bits, so code executing from the ROM crashes and the image is corrupted. The shipped monitor is safe by construction (firmware/monitor/monitor.asm:58-61: first instruction at $F000 is BR eprom to $F003, which clears FORCE-ROM before any store), but nothing in hardware enforces that.

Bench: with the bus tester, reset, write a byte to $0010 with -MEM-WR, then read $F010 through the ROM: it changed (use a scratch EEPROM). With a meter, -WE on IC13.27 follows bus pin B24 one-for-one.

M3 — MED: TMP registers latch on the leading edge of -TMP-REG-LDn; correct only under a microcode ordering rule

  • IC26/IC27 (74LS374) CLK(11) = N$33 = IC14.2 = NOT(-TMP-REG-LD0); IC28/IC29 CLK = N$47 = IC14.12 = NOT(-TMP-REG-LD1). A 374 clocks on the rising edge of CLK, i.e. ~10 ns after the falling (leading) edge of the bus strobe, so whatever is on DATA0-15 at the start of the strobe is captured.
  • Every register-card source needs far longer than 10 ns to appear on the bus after its own strobe (LS32 -> two LS139 -> LS32 -> 244 enable -> CD4077 (~100 ns, see R1) -> LS32 -> 245 enable); a memory read needs ~150 ns.

The microcode compensates by asserting the data source one line before the LD strobe (memory.c:24-28, accumulator.c:343-349, branch.c:624-631) — the hardware silently depends on that. branch.c:14-22 (moveRegtoTmp) asserts -REG-FUNC-RD/-REG-RD-LO/-REG-RD-HI and -TMP-REG-LD0 in the same line and would latch stale data; it currently has no callers, so the fault is latent.

Bench: bus tester: put a value on DATA, assert -TMP-REG-LD0 low, change DATA while it is still low, release, read back with -TMP-REG-RD0: the first value is returned.

M4 — MED (configuration): low RAM cannot be removed from the map; conflicts with the bring-up cards and with FORCE-ROM

  • -LO-RAM = IC10.3 (74ALS00) = NAND(N$13 = NOT BADDR15 [IC14.4], N$12 = NOT -VMA [IC14.10]). No jumper, no block decode: $0000-$7FFF is always RAM when -VMA is low.
  • The Mem Switch card decodes $0000-$000F and the Mem Register card $0010-$001F (comparators IC4 on each, see B1), both driving DATA0-7 on -MEM-RD.

Failure scenario: if the memory card is on the bus together with either bring-up card (MACHINE.md says they were re-fitted 2026-09-21), every read of $0000-$001F has two drivers on DATA0-7: the 62256 through IC5 (74LS245, DIR = -MEM-RD, G = -VMA) and the bring-up card's 74LS244/374 outputs. After reset it is three, because FORCE-ROM maps the EEPROM onto $0000 as well. Bench: meter on any DATA0-7 line during a read of $0000 shows an intermediate level (~1-2 V) if two drivers disagree.

M5 — LOW: control inputs rely on floating-high when -BUS-EN is high

-MEM-WR, -MEM-RD, -VMA, -TMP-REG-RD0/1, -TMP-REG-LD0/1 come from 74LS374 outputs on the sequencer whose OC pins are -BUS-EN (sequencer IC7/12/14/28 pin 1), so all of them are tri-state until the sequencer-memory card reports READY (~54 s after power-up) and whenever the bus is handed over. No card and not the backplane has a pull-up on any of them (checked on all five cards plus io and sequencer). LS inputs float high so nothing is selected, but the levels are undefined and the 62256/28C64 -WE depends on it. Bench: meter on B24 with the sequencer un-READY reads ~1.5-1.9 V, not a logic high.

M6 — LOW: unused 74LS04 inputs left open

IC14 pins 5 and 9 have no net at all in the PCB (pads unconnected), so the mechanical pass, which looks for single-node nets, could not see them. IC3 and IC12B unused inputs are tied off correctly.

M7 — LOW: IC7 (74LS138) G2B needs JP1 fitted

IC7.5 = N$24 = JP1 pin 2 only; JP1 pin 1 = -BUS-EN, pin 3 = GND. With no jumper the pin floats high and the whole high-32K decode (RAM blocks and ROM) is dead. Note MACHINE.md's "jumper wire from IC7 pin 4, unconnected": pin 4 is G2A = -VMA; if that wire ever replaces the track, pin 4 floats and the same symptom appears.

M8 — LOW: DATA0-15 pull-down networks RN5/RN6 have no value in the design files

RN5/RN6 (RNX8, pin 1 = GND) are pull-downs on the bus DATA0-15. Their value is empty in both .sch and .brd. Below ~4.7 k they would exceed the LS IOH budget of every driver on the bus; MACHINE.md's observation that an undecoded read "returns the last value left on the bus" suggests they are either high-value or not fitted. Confirm with a meter (resistance from A19 to GND with the card out).

Checked, no issue (memory card)

  • 62256 -OE = N$3 = -MEM-RD via two inverters; -WE = -MEM-WR; never both low; IC5 DIR = -MEM-RD so the 245 receives during writes and drives during reads. Only a ~20-30 ns BDATA overlap at the end of a read (245 turns round before the RAM's -OE has released) — normal for LS designs.
  • Write strobe ordering: main.c:236-249 asserts ADDR-ID + -VMA + source in one line and -MEM-WR in the next, so address decode (IC8/9 244s, IC11 157, IC7 138, IC4/IC18 NAND8s) has a full step to settle before -WE falls, and the source is held one line after -MEM-WR rises: no cross-block write from decode glitches.
  • Block jumpers: IC7 Y0-Y7 -> U$1 middle row; top row -> IC4 (74ALS30, pull-ups RN8) -> IC6.10 -HI-RAM; bottom row -> IC18 (pull-ups RN7) -> IC6.12 -ROM-CS. Unjumpered blocks are pulled inactive. ROM A12 = BADDR12 so both halves of the 8K are reachable when $E and $F are jumpered down.
  • FORCE-ROM uses the raw ADDR15 (IC10.4) and not BADDR15, so the 157 feedback loop cannot hold it set.
  • Reset polarity: IC12A PRE = -RESET (active low) matches the bus.
  • Power-on state of FORCE-ROM is defined only by -RESET: see S1.

Last updated 26 October 2026