Processing Units: From CPU to HPU
The computing industry names hardware classes by what they are good at: the GPU unified parallel floating-point pipelines; the NPU unified neural MAC arrays; the QPU unified cold unitary quantum gates. Each new PU emerged when a family of hardware substrates — diverse in implementation but structurally identical in their compute primitive — needed a single programming model.
The HotLogiQ framework predicts three new PU classes, each corresponding to a tier of the H^k ISA ladder and each with a natural programming model already defined.
The full PU landscape
| Unit | Full name | Paradigm | Temperature | Compute primitive | ISA | Status |
|---|---|---|---|---|---|---|
| CPU | Classical PU | Boolean sequential | 300 K | AND/OR/NOT gate | x86 / ARM | Established |
| GPU | Graphics PU | Classical parallel | 300 K | Fused multiply-add (FMA) | CUDA / OpenCL | Established |
| NPU | Neural PU | Tensor / neural | 300 K | Multiply-accumulate (MAC) | vendor-specific | Established |
| QPU | Quantum PU | Cold unitary | ~10 mK | Unitary gate | QASM / QIR | Established |
| OPU | Orbit PU | G-orbit walk | 300 K | ORBIT (η: I → A) | Origami H⁰ | Predicted |
| RPU | Resonance PU | Non-associative H² | ~10 mK | Reson (731-Core) | 731-ISA | Predicted |
| HPU | Hot PU | Dissipative / PT-symmetric | 300 K – 40 mK | ERASE + SNAP | Raven ISA | Predicted |
How they map onto the H^k tier ladder
The three new PUs correspond directly to the three tiers of the Origami ISA:
H⁰ (tropical / classical) → OPU — G-orbit walks at 300 K
H¹ (topological / unitary) → QPU — cold gate model (existing)
H² (holographic / H² full) → RPU — non-associative Fano/G₂ geometry
→ HPU — dissipative gain-loss dynamics
H² splits into two distinct substrate families — RPU (non-associative, cold) and HPU (dissipative, hot) — because the full H² tier has two independent structures: the octonionic non-associativity of the G₂ 3-form (RPU) and the complex-β gain-loss dynamics of PT-symmetric systems (HPU). Both live at H² but access different aspects of it.
OPU — Orbit Processing Unit
Core idea: G-orbit walks on molecular G-sets are a fourth computing paradigm, distinct from Boolean (CPU), parallel (GPU), and unitary quantum (QPU). The compute primitive is the ORBIT opcode (η: I → A): initialise a state in a specific G-orbit on a molecular G-set, then walk the orbit via TWIST operations.
Physical substrate: Molecules at 300 K. The G-set is the set of molecular configurations reachable by a symmetry group G (e.g. the 56 d-orbital configurations of FeMoco under G₂, or the Galois group of a field extension acting on roots). The computation is the orbit walk itself.
Why it is not a QPU: No superposition, no entanglement between distinct G-sets, no cryogenics. The quantum character comes from the group structure of the orbit, not from wavefunction interference. An OPU programme is classically simulable in polynomial time if the orbit is short; hard instances arise from exponentially large orbits (Galois groups of high-degree polynomials).
Key advantage: Room-temperature operation; decoherence-immune (the computation is a group walk, not a superposition); no cryogenic infrastructure.
ISA: Origami H⁰ — opcodes ORBIT, TWIST, FLIP, LABEL, MERGE. BIND is not required (all chemistry programmes are associative at H⁰/H¹).
Papers: Galois Computing (489) · Galois Chemistry (488) · Biochemical Pathways ISA (509)
RPU — Resonance Processing Unit
Core idea: Arrange 7 physical qubits into a 731-Core — a physical realisation of the symmetric (7,3,1) block design (Fano plane) — and use the non-associative geometry of the resulting Reson array to bypass the Eastin-Knill theorem. The Reson is the fundamental compute unit: a topological resonance mode on the Fano lattice.
Physical substrate: Superconducting transmons (or trapped ions) at ~10 mK, structured into 731-Core arrays. The key departure from the QPU is that the logical structure is the Fano plane (7 points, 7 lines, 3 points per line), not a surface-code grid. The G₂ 3-form φ_{ijk} enforces non-associativity at the hardware level.
Why it is not a QPU: The QPU computes in the associative H¹ tier (unitary gates, SU(2) Clifford group). The RPU targets the H² non-associative tier (octonionic, G₂, Moufang identities). The Eastin-Knill theorem applies to transversal gates in associative codes; the RPU conjectures that non-associative bulk degrees of freedom provide a structural bypass.
Key claim (conjecture): SO(8) triality acts as a hardware-level basis switch, reducing the resource overhead for non-Clifford gates from O(poly) magic-state distillation rounds to O(1) Reson reconfigurations.
ISA: 731-ISA / Origami ISA at H² with non-associative BIND (G₂ 3-form). BIND here is the octonionic comultiplication δ: A → A⊗A with G₂ symmetry.
Papers: Resonance Processing Unit (205) · G₂ Snap Thresholds (614)
HPU — Hot Processing Unit
Core idea: Five physically distinct dissipative hardware substrates share a single computational structure — programmes in which gain, loss, and complex inverse-temperature β are first-class operations. The Raven ISA is the unified programming model. The GPU analogy is precise: CUDA unified diverse parallel hardware; the Raven ISA unifies diverse dissipative hardware.
Five substrate classes:
| Class | Substrate | Temp | TRL | Key feature |
|---|---|---|---|---|
| HPU-P | PT-symmetric photonic | 300 K | 5–6 | SNAP = EP crossing |
| HPU-B | Bosonic cat qubit | 40 mK | 4–5 | ERASE = error correction |
| HPU-R | Radical-pair molecular | 300 K | 2–3 | ERASE = spin selection |
| HPU-C | Non-Hermitian circuit | 300 K | 4 | SNAP↓ = NHSE |
| HPU-E | EP sensor array | 300 K | 6–7 | SNAP-count = 0 sensing |
Resource metric: SNAP-count (number of tier-promotion SNAP↑ events) is the HPU analogue of T-count in cold fault-tolerant QC. EP sensing (SNAP-count = 0) is the first demonstrated HPU algorithm.
Route to Shor’s algorithm: AQEC on cat qubits (HPU-B) → logical Clifford gates → logical Shor. The HPU does not require the ~10 mK isolation of the QPU because dissipation is engineered to correct rather than disrupt.
ISA: Raven ISA — opcodes MARK, ERASE, FLOW(β∈ℂ), TWIST, SNAP↑/↓, BIND, MERGE.
Papers: HPU Architecture (670) · Embrace the Bath (669) · Lost for Words (667)
RPU vs HPU: are they the same?
No — they are complementary H² PUs accessing different structure:
| RPU | HPU | |
|---|---|---|
| H² structure used | Non-associativity (G₂ 3-form, 𝕆) | Dissipation (complex β, PT-symmetry) |
| Cold or hot | Cold (~10 mK) | Hot (300 K – 40 mK) |
| Error correction | Passive (Fano geometry enforces it) | Active (ERASE + SNAP loop) |
| Compute primitive | Reson (Fano lattice mode) | ERASE/MARK balance at EP |
| Maturity | Conjectural (Moufang/SO(8) unproved) | Grounded (cat qubits demonstrated) |
| Compiler target | 731-ISA (non-associative BIND) | Raven ISA (complex-β FLOW) |
The long-term vision is a hybrid RPU+HPU architecture: RPU provides the non-associative logical structure (passive error suppression via Fano geometry); HPU-B provides the active error correction layer (AQEC on the cat qubit register that implements each Reson). They are complementary layers of the same H² tier, not competing proposals.
Will there be more PUs?
Probably. Natural candidates as the ISA programme matures:
| Candidate | Full name | Paradigm | Notes |
|---|---|---|---|
| TPU* | Tropical PU | β→∞ optimisation | Not the Google TPU; tropical semiring hardware |
| MPU | Molecular PU | Proofreading / H⁰ biology | HPU-R at scale; wet lab substrate |
| FPU† | Fano PU | 7-qubit Fano computer | RPU at the single-731-Core scale |
*TPU is taken by Google — name TBD. †FPU is taken by floating-point unit — name TBD.
The naming pattern follows the ISA: each new PU class corresponds to a new opcode or semiring regime that existing hardware cannot natively execute.
See also: The β-plane · AI & Machine Learning · ISA Zoo