Thermo3inside the Eurorack question

Four oscillators are the easy part.

What would it take to build a Eurorack instrument that behaves like Thermo3?

The answer begins with a feedback network that listens to itself. Four voices exchange sound; the instrument senses their activity, remembers interventions and changes the paths that sound takes next.

Four voice feedback network Four oscillator voices surround a four by four matrix. Amber arrows show sound circulating between voices, while green branches mark audible outputs. 01020304 4 × 4 routing sound returns to the voices

A conceptual map: each voice has a path to the listener and a separate path back into the network.

First decide what “equivalent” means.

A modular patch can travel through similar sonic territory without copying the way Thermo3 makes its decisions. These are three different ambitions.

Similar musical territory

Evolving drones, beating, unstable textures and hands-on intervention. A well-designed rack can reach this territory with fewer exact correspondences.

Comparable architecture

Four sound-making and receiving voices, adjustable links between them, separate audible and return paths, sensing and independent movement.

Full behavioural match

The same adaptive routing, captured gestures, relative activity memory, per-voice heat state, clones and coordinated recall. This requires a custom design and testing.

Thermo3’s defining feature is the relationship between circulation, sensing, memory and intervention.

Follow one voice through the loop.

Imagine one oscillator receiving sound from the other three. It adds its own source, processes the combination, and then takes two different routes.

Thermo3 signal flow through one voice Local source and routed feedback combine, then pass through delay, all-pass and filter. The filter output splits into audible output and a feedback branch with tap delay, level regulation, selected Transfer, Thermal, sensing and matrix return. Local sourceMatrix inputSumDelay · all-pass · filterAudible voicepanner / individual output Tap delayLevel controlTransferThermalRMS / reaction → 4 × 4 matrix The split happens here, before Transfer The return changes what happens next
Transfer and Thermal affect the returning network signal. Their sound is heard as the voices react on subsequent passes, rather than as final effects on the dry stereo output.

The audible path

The filtered voice goes to the stereo panner or its individual output. Four individual oscillator outputs can be used without changing the feedback design.

The returning path

A separate branch is delayed, regulated, transformed and heated. The instrument measures its activity and sends it back through the routing matrix.

Thermo3 interface showing four oscillator voices, feedback matrix, Transfer and Injection controls
Thermo3 itself: four voices, a visible routing matrix and independent intervention controls in one instrument.

What would the rack need?

The module names here are examples of useful roles, not a verified patch or an endorsement of a particular buying list. The real design problem is keeping enough independent paths available at once.

Four flexible sources

One per voice. Generate 3 illustrates frequency, phase, amplitude and harmonic control; it is not a copy of Thermo3’s source morphing.

Four receiving filters

Each voice needs its own filtering after local and network audio combine. Filter 8 illustrates multimode control, with a different topology from the software.

A four-by-four matrix

A 4ms VCA Matrix offers sixteen controllable routes. Cross-voice connections can be selected while diagonal routes stay muted.

Four return processors

Delay, level regulation, nonlinear Transfer and Thermal-like behaviour belong in the return paths, with simultaneous channels.

Activity sensing

Followers, VCAs, comparators, scaling and smoothing let the instrument respond to how the four voices behave.

GEN movement

Several independent rates, random distributions and destinations need clocks, random voltages, function generators and utilities.

Audible mix and safety

Four-voice mixing or individual outs, panning where wanted, line output and deliberate gain protection.

State and recall

Preset-like recall needs a strategy for voltages, routing, hidden state and physical patch cables.

122 HP is only the starting footprint. Four 12 HP Generate 3 modules, four 12 HP Filter 8 modules and one 26 HP VCA Matrix add up to 122 HP. Return processing, sensing, modulation, mixing, case space and power still have to be allocated.

The hard part is the instrument’s behaviour.

Modular hardware can produce related sounds. These operations ask the hardware to keep track of time, activity and the consequences of a performer’s action.

Injection is an intervention in a circulating sound network A performer gesture or capture enters an Injection response, influenced by measured voice activity and memory. The response changes the circulating network and the four voices react. Gesture / captureActivity + memoryInjection responseplus optional LayersCirculating audioand routingFour voices The voices’ next reaction feeds the loop
Injection changes the system that the sound moves through. Its value is in the reaction, not merely in an added effect.

Injection: a gesture with consequences

Strike introduces an excitation; Quench drains circulation while local sources remain. Reroute shifts connections according to activity. Threshold Dust makes sparse activity-dependent events.

Imprint stores the relative activity of the four voices. Afterimage captures raw audio material and returns it to the network; it replaces the earlier standalone Habituation Response. Habituate and Dwell remain modifiers that shape how a Response develops.

To approximate these operations in hardware, a gate or button is only the beginning. The rack also needs measurement, memory, return paths and several control channels working together.

Transfer and Thermal: history in the return

Clip, atan, rectify and fold have recognizable modular relatives. Hysteresis, Excitable and Bistable rely on state and response rules; a distortion module alone cannot establish the same behaviour.

Spectral Drift shifts frequency by small offsets that accumulate on each return. Pitch transposition is a different process. Flow, Opposition and Adaptive need their own direction behaviour per voice.

Thermal follows Transfer. It senses activity, accumulates and loses heat over time, and changes colour, filtering, limited pitch movement and optional noise. The four voices each carry their own heat state.

Clones double more than the oscillator count. Reproducing four clone voices means additional source and processing paths, controlled entry into the network and enough modulation and sensing capacity to keep all eight paths musically useful.

A representative hardware budget.

These figures describe one plausible four-voice architecture. They were gathered for the original assessment on 17 October 2026 and are not live quotations or a tested bill of materials.

£7k–£10kRounded estimate for four primary voices and substantial control/return processing.
£10k–£15k+Rounded estimate when four extra clone-processing paths are planned.

The detailed calculation for the four-voice version was £6,829–£9,529: £3,729 in specifically priced modules plus £3,100–£5,800 of planning allowances. Clone paths add a provisional £3,000–£5,000. The two disting NT units in the priced line are a capacity placeholder; concurrent DSP performance has not been verified.

Show the original budget line by line
Item or allocationBasisBudget
4 × Generate 34 × £299£1,196
4 × Filter 84 × £255£1,020
4ms VCA Matrix1 × £315£315
Provisional DSP: 2 × disting NT2 × £599£1,198
VCAs, followers, envelopes and utilitiesPlanning allowance£1,000–£1,800
Random modulation and function generatorsPlanning allowance£700–£1,400
Stereo mixing, panning and outputPlanning allowance£600–£1,200
Powered cases, cables and accessoriesPlanning allowance£800–£1,400
TotalFour primary voices£6,829–£9,529

Custom firmware and engineering, assembly, patching, commissioning, comparison tests, delivery, computer, DAW, interface and monitoring are outside this estimate. A different module selection, used hardware or a custom DSP design would change it.

Prototype the loop before building the rack.

A module list alone cannot prove that the behaviours can run together. Build and listen in stages, adding complexity only after each layer works.

  1. One voice, two branchesCombine a source with incoming audio, keep audible and feedback paths separate, and establish reliable gain control and silence.
  2. Two voices in conversationMake distinct sources exchange sound audibly. Test Quench on the return without removing the local sound.
  3. Four independent routesVerify the matrix and activity sensing across the full intended level range.
  4. Intervention and memoryTest relative Imprint capture, Afterimage audio return, adaptive Reroute and the Response modifiers as separate operations.
  5. Everything at onceMeasure whether delays, stateful Transfer, frequency shift, Thermal, Layers and GEN movement fit the processing/control budget.
  6. Clones and performanceBring extra paths in cleanly, then decide how a musician will patch, control and recall the complete instrument.

Sources and limits

The module roles and October 2026 UK prices come from the original Thermo3 and Eurorack: Architecture and Cost Assessment. Thermo3’s signal-flow and current Injection terminology were checked against its project and consolidated manual. Manufacturer pages describe capabilities; retailer listings underpinned the dated prices. No complete rack has been assembled or benchmarked for this comparison, and no sound-quality ranking follows from it.