First Dynamical QCD Production on Consumer GPU

Date: March 9, 2026 Status: Complete (all 17 β points finished, 1,071 trajectories). Chuna Papers 43-45: 44/44 overnight checks pass (v0.6.24). Dynamical N_f=4 ext 3/3 complete. 🪸🌊 coralReef sovereign compilation: 44/46 shaders, full GpuBackend impl. Spring: 🔥♨️ hotSpring (v0.6.24) Hardware: RTX 3090 (GPU) + BrainChip AKD1000 (NPU) + Titan V (DRM testing) License: AGPL-3.0-only

Summary

First dynamical fermion lattice QCD production scan on a consumer GPU. An 8⁴ lattice with staggered quarks (N_f = 1, m = 0.1) scanned across 17 β values steered by a 14-head neuromorphic coprocessor. The NPU expanded a 4-point seed scan to 17 points in real time, systematically mapping the confined → deconfined crossover. The crossover is smooth (no first-order jump), confirming the expected qualitative change from quenched QCD. β_c has shifted downward from 5.692 (quenched) to approximately 5.0–5.5 (dynamical, 1 flavor).

1. Background: Why This Run Matters

⛺📄 baseCamp 07 established quenched SU(3) lattice QCD on consumer hardware: two production scans at 32⁴ showing β_c = 5.69 (matching literature to three significant figures), DF64 hybrid arithmetic (2× speedup), and NPU adaptive steering (2.5× more useful statistics at same wall time).

Quenched QCD ignores quarks. The gluon field evolves alone — no virtual quark-antiquark pairs, no fermion backreaction. The deconfinement transition is first-order (a sharp discontinuity in thermodynamic quantities). This is computationally cheaper but physically incomplete.

Dynamical QCD includes the fermion determinant, adding a Conjugate Gradient (CG) solver that dominates the computational cost. The transition softens from first-order to a smooth crossover. The critical coupling shifts. The physics is qualitatively different.

This run is the first dynamical production scan on the biomeGate system and, to our knowledge, the first NPU-steered dynamical QCD scan of any kind.

2. NPU Adaptive Steering: How the Scan Grew from 4 to 17 Points

Initial seed

The run was launched with 4 seed β values: 5.0, 5.5, 5.69, 6.0.

NPU steering decisions (chronological)

The NPU evaluated each completed β point and chose where to insert the next scan point. The full steering trace:

The NPU locked onto β_c ≈ 5.50 after seeing the first 3 points and never revised this estimate. It then spent most of its steering budget mapping the confined side of the transition — inserting 11 points below β = 5.0 where the crossover to confinement occurs. This is a notable difference from the quenched runs, where the NPU focused on the transition region (β ≈ 5.4–5.8). The dynamical crossover is broader and the NPU correctly identified that the interesting physics extends much further into the confined regime.

Steering overhead

3. Production Results

Per-β data table (all 17 completed)

Sorted by β. The “Order” column shows when each point was evaluated — note how the NPU jumped between high and low β rather than scanning linearly (see scan trajectory diagram below).

Plaquette curve (seeds vs NPU-inserted)

 ⟨P⟩
 0.60 │                                                         o  ← 6.13  NPU #9
 0.59 │                                                     o      ← 6.07  NPU #7
 0.58 │                                                 S          ← 6.00  SEED #3
 0.55 │                                            S               ← 5.69  SEED #1
 0.53 │                                       S                    ← 5.50  SEED #2
 0.40 │                                  S                         ← 5.00  SEED #4
 0.39 │                             o                              ← 4.93  NPU #5
 0.38 │                         o                                  ← 4.86  NPU #6
 0.37 │                     o                                      ← 4.79  NPU #8
 0.36 │                 o                                          ← 4.73  NPU #10
 0.35 │             o                                              ← 4.67  NPU #11
 0.35 │         o                                                  ← 4.61  NPU #12
 0.34 │      o o                                                   ← 4.55  NPU #13
 0.33 │   o o                                                      ← 4.44  NPU #15
 0.32 │  o                                                         ← 4.39  NPU #16
 0.32 │ o                                                          ← 4.34  NPU #17
      ┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──┼──
      4.3 4.5 4.6 4.7 4.8 4.9 5.0       5.5  5.7  6.0 6.1  β

      S = seed point (human-selected)     o = NPU-inserted

NPU scan trajectory (evaluation order)

The NPU did not scan linearly. It bracketed the transition region, alternating between the low-β confined side and the high-β deconfined tail. This diagram shows the order each β was evaluated (read left to right), with arrows showing where the NPU jumped:

  Evaluation order →

  Step:  1     2     3     4     5     6     7     8     9    10    11    12    13    14    15    16    17
  β:   5.69  5.50  6.00  5.00  4.93  4.86  6.07  4.79  6.13  4.73  4.67  4.61  4.55  4.50  4.44  4.39  4.34
  Src:  [S1]  [S2]  [S3]  [S4]  NPU   NPU  NPU   NPU  NPU   NPU   NPU   NPU   NPU   NPU   NPU   NPU   NPU
        │     │     │     │     │     │     │     │     │
        └──┬──┘     │     │     │     │     │     │     │
       start at     │     │     │     │     │     │     │
       transition   │     │     │     │     │     │     │
                    │     │     └──┬──┘     └──┬──┘     │
                    │     │    low-β fill    jump to    high-β
                    │     │                 high-β      tail
                    └─────┴───────────────────────────────── then systematic downward sweep ───→

  β (number line, showing jump pattern):

  4.3   4.5   4.7   4.9   5.0       5.5   5.7   6.0   6.1
  ├──────┼─────┼─────┼─────┤         ├─────┤─────┤─────┤
  17←16←15←14←13←12←11←10  ↑  ←8  ←6  ↑     ↑     ↑  ←7  ←9
                             5         4     2     1     3

  Read: The NPU started at seed 5.69 (#1), jumped to 5.50 (#2),
  then 6.00 (#3), then 5.00 (#4). After these 4 seeds, it inserted
  4.93 (#5) and 4.86 (#6), then jumped up to 6.07 (#7) to balance,
  back down to 4.79 (#8), up to 6.13 (#9), then swept systematically
  downward: 4.73, 4.67, 4.61, 4.55, 4.50, 4.44, 4.39, 4.34.

Key observation: The NPU’s “bracket and fill” strategy is visible in the jump pattern. After the 4 seeds established the range, it alternated high/low insertions (steps 5-9) to bracket the crossover from both sides before committing to a systematic downward sweep (steps 10-17). This is exactly how a physicist would explore an unknown phase diagram — coarse bracketing first, then fine filling. The NPU learned this strategy from the quenched training data without being explicitly programmed to bracket.

The plaquette rises monotonically and smoothly from ⟨P⟩ = 0.316 at β = 4.34 to ⟨P⟩ = 0.596 at β = 6.13. No discontinuity — this is the smooth crossover expected for dynamical fermions, in contrast with the quenched first-order transition.

The steepest gradient is between β ≈ 5.0 and β ≈ 5.5 (ΔP/Δβ ≈ 0.23), consistent with the crossover region. This is well below the quenched β_c = 5.692, confirming the expected downward shift from fermion backreaction.

4. Key Physics Findings

4.1 The crossover is smooth

The quenched deconfinement transition at β_c = 5.692 is first-order — the plaquette jumps discontinuously. The susceptibility χ is sharp and tall (χ ~ 40–53 in the 32⁴ quenched runs).

The dynamical run shows no discontinuity at any β. The plaquette varies smoothly and the susceptibility is small and broad (χ < 0.25 everywhere). This is the expected crossover behavior: dynamical quarks screen the gluon self-interaction, washing out the first-order transition.

4.2 β_c has shifted downward

In quenched SU(3), the deconfinement transition occurs at β_c = 5.692 (known from decades of lattice calculations). With 1 flavor of staggered quarks at m = 0.1, the steepest plaquette gradient sits between β ≈ 5.0 and β ≈ 5.5. The NPU’s β_c estimate of 5.50 is consistent with this. Fermion backreaction adds attractive forces at the confinement scale, lowering the critical coupling.

4.3 CG cost varies systematically with β

CG iterations decrease by ~19% from strong to weak coupling. This reflects the improving condition number of the Dirac operator: at weak coupling, gauge fluctuations are smaller (less “disorder” in the Anderson analogy), the lowest eigenvalue is larger, and the matrix is easier to invert. The acceptance rate improves correspondingly from ~50% to ~80%.

This systematic CG–β correlation is exactly what the Anderson proxy pipeline (Exp 026) is designed to predict cheaply.

4.4 The Polyakov loop is noisy but present

The Polyakov loop magnitude |L| ≈ 0.29 across all β values. On an 8⁴ lattice, the Polyakov loop has large finite-volume fluctuations and is not a clean order parameter. At 32⁴, we expect |L| to show clear separation between confined (|L| → 0) and deconfined (|L| → finite) phases, as it did in the quenched runs.

5. Comparison: Quenched vs Dynamical

The dynamical run has fewer total trajectories but more β points, because the NPU correctly identified that the broad crossover requires denser sampling over a wider β range. The quenched transition is sharp and localized — 10 points suffice. The dynamical crossover spans Δβ ≈ 1.5 — the NPU mapped it with 17 points.

6. NPU Performance

CG prediction accuracy

The NPU’s CG estimates varied widely:

The CG estimates are systematically low because the ESN was trained on quenched data where there is no CG solver. The NPU has no prior dynamical training data — this run IS the first training set. The Exp 026 proxy pipeline (4D Anderson + Wegner) will provide physics-informed CG predictions that should dramatically improve this.

What worked well

• β_c estimation: Locked to 5.50 after 3 points and stayed stable. This is reasonable for 1-flavor dynamical fermions.
• Adaptive steering: Expanded 4 → 17 points, systematically mapping the full β range. Correctly identified that the crossover extends far into the confined region.
• Phase classification: Correctly labeled all β < 5.5 as “confined” and β = 5.69 as “transition.”
• Anomaly detection: Flagged 5 anomalies per β point — likely the first few trajectories after thermalization that haven’t fully equilibrated. Consistent behavior suggests a real pattern, not noise.

What needs improvement

• CG prediction: Needs dynamical training data (this run provides it) and physics proxy input (Exp 026).
• Parameter suggestion: The NPU consistently suggested smaller dt and larger n_md than what was used (e.g., dt=0.001 vs actual dt=0.01). The suggestions were more conservative but the defaults worked, so the NPU was being cautious without data.

7. What This Means for Scale-Up

The 8⁴ run validated:

1. The fermion force is correct — acceptance is 60%, ΔH is O(1), the plaquette curve is physical. The bug fix from Exp 024 (momentum kick sign error) is confirmed stable over 1,000+ trajectories.

2. NPU steering works for dynamical QCD — the scan expanded sensibly, β_c estimation is stable, phase classification is correct.

3. CG prediction needs physics proxies — the ESN alone (without Anderson/Wegner training data) cannot predict CG iterations for a new physics regime. This is the primary motivation for Exp 026.

4. The crossover is broader than expected — the NPU inserted 13 additional points and the physics hasn’t plateaued at the low end. A 32⁴ production run should plan for β range 4.0–6.5 with 20+ points.

Scale-up roadmap (updated from Exp 025)

8. Energy Context

Observed thermals

GPU temperature during this run was significantly lower than the quenched 32⁴ runs:

The 8⁴ lattice uses 0.06% of VRAM and ~20% of shader cores. Most of the 3090’s transistors are idle. This means the CG solver, despite being the dominant cost, is not GPU-limited — it’s algorithmically limited by the number of iterations, not by the available FLOPS.

Scaling to 32⁴ dynamical will bring GPU utilization and thermal output back to quenched-run levels. See Exp 027 for full energy tracking specifications.

9. Future Directions

Immediate (before next production run)

• Exp 026: Run 4D Anderson + Wegner block proxy pipeline to generate physics-informed CG training data for the NPU.
• Exp 025A: 16⁴ single-β validation to measure real CG scaling.
• Exp 027: Instrument energy tracking in all production binaries.

Medium-term

• 2+1 flavor: Add a second pseudofermion field for the strange quark, matching the physical QCD configuration. Doubles CG cost per trajectory.
• 32⁴ dynamical production: Full-volume scan with NPU steering trained on Exp 024 + Exp 026 data.

Connection to other baseCamp papers

• ⛺📄 baseCamp 01 (Anderson QS): The CG–disorder correlation observed here directly validates the Anderson localization framework. Gauge fluctuations at strong coupling (high plaquette variance = high effective disorder) produce harder CG solves, exactly as Anderson predicts.
• ⛺📄 baseCamp 07 (WDM/QCD): This run extends paper 07 from quenched to dynamical. The DF64 arithmetic, NPU steering, and vendor-agnostic shader stack carry over unchanged.
• ⛺📄 baseCamp 04 (Sentinels): The multi-head NPU architecture demonstrated here (14 heads, real-time steering) is the same pattern used for environmental biosensing — cheap inference guiding expensive measurement.

Addendum: NPU as Parameter Controller (Exp 031, 2026-03-01)

Exp 030 revealed that the NPU’s parameter suggestions (dt, n_md) were being received but never applied. The auto_dt formula over-penalized mass (mass_scale.sqrt() turned dt=0.01 into dt=0.0032 for mass=0.1), producing 97.5% acceptance — far above the 60-80% sweet spot and wasting ~2x CG iterations per useful trajectory.

Exp 031 makes the NPU the actual controller of HMC parameters:

Mid-beta feedback loop fires every 10 measurement trajectories: if acceptance

85%, dt bumps 15%; if < 50%, dt drops 15%. The dt_used and n_md_used fields in BetaResult enable post-hoc analysis of how the NPU adapts parameters across the phase curve. Safety clamps: dt ∈ [0.001, 0.02], n_md ∈ [20, 500]. A --no-npu-control flag reverts to the old print-only behavior.

This closes the gap between what the NPU knows and what the NPU controls — the brain architecture now has a complete feedback loop from measurement to parameter adjustment, with the Titan V pre-motor receiving the NPU-adapted dt for the next beta point.

Addendum: Deep Debt Resolution (v0.6.18, 2026-03-06)

🔥♨️ hotSpring v0.6.18 completed a comprehensive technical debt audit (Exp 041): Clippy 0 warnings (pedantic+nursery), file-size compliance (<1000 lines), unwrap/expect removal from production sites, SPDX 100% AGPL-3.0-only. Brain B2 (memory pressure) and D1 (force anomaly) evolved from placeholder to real runtime estimates. 685 lib tests pass. See hotSpring/experiments/041_DEEP_DEBT_RESOLUTION_AUDIT.md.

Data Files

References

• A. Bazavov et al. [HotQCD]. “Equation of state in (2+1)-flavor QCD.” Phys. Rev. D 90, 094503 (2014).
• T. G. Kovács and F. Pittler. “Anderson Localization in Quark-Gluon Plasma.” Phys. Rev. Lett. 105, 192001 (2010).
• M. Giordano, T. G. Kovács, F. Pittler. “Dirac mode localization in QCD near the crossover temperature.” arXiv:2602.10921 (2026).
• B. Svetitsky and L. G. Yaffe. “Critical behavior at finite-temperature confinement transitions.” Nucl. Phys. B 210, 423 (1982).
• F. Wegner. “Disordered system with n orbitals per site: n = ∞ limit.” Phys. Rev. B 19, 783 (1979).