hotSpring — Computational Plasma Physics, Lattice QCD, Spectral Theory

Dense plasmas, nuclear structure, lattice QCD — 697+ tests on consumer GPUs for $0.044. Paper parity on a $600 card.

Domain

Dense plasmas, nuclear structure, molecular dynamics, lattice QCD, spectral theory, neuromorphic computing.

Repository: syntheticChemistry/hotSpring

The Science Story

hotSpring is the primary GPU science driver — the spring that proves barraCuda can do first-principles computational physics on consumer hardware. It is the most mature spring because physics has the least room to hide: 0.000% energy drift or the simulation is wrong.

Headline Results

  • Sarkas Yukawa MD at paper parity (N=10,000, 80k steps) on a $600 RTX 4070 for $0.044 in electricity
  • Full AME2020 nuclear dataset (2,042 nuclei — 39× the published paper) on a single consumer GPU
  • Lattice QCD β-scans (32⁴, 12 temperatures) resolving the deconfinement transition on a $500 RTX 3090 for $0.58
  • DF64 delivers 3.24 TFLOPS of double precision on FP32 cores
  • Phase 0 discovered and fixed 5 silent bugs in the upstream Sarkas code

Validation Phases

PhaseKey Result
A–E (MD)Python → Rust → GPU → f64 → paper parity. 0.000% energy drift. $0.044 electricity
F (Nuclear EOS)2,042 nuclei AME2020 on consumer GPU. 478× speedup, 44.8× energy reduction
Lattice QCDSU(3) HMC + dynamical fermions. 32⁴ β-scan, deconfinement at β=5.69
SpectralAnderson localization (1D/2D/3D), Hofstadter butterfly, Lanczos eigensolver
NPU10 SDK assumptions overturned. ESN streaming at 2.8μs/step

Researchers Reproduced

ResearcherDepartmentDomain
Michael MurilloCMSE, MSUDense plasmas, WDM, molecular dynamics
Alexei BazavovCMSE + Physics, MSULattice QCD, thermodynamics
Ilya KachkovskiyMath, MSUSpectral theory, Anderson localization
Rika AndersonBiology, CarletonPangenomics (cross-spring)

What the Constraint Revealed

Eliminating CUDA forced Vulkan, which exposed SHADER_F64 on consumer GPUs. Eliminating vendor compilers forced coralReef, which now compiles 93/93 cross-spring WGSL shaders to native GPU binaries. A $300 Akida NPU runs ESN inference at 2.8μs/step — 1,000× faster than GPU for streaming workloads, 9,017× less energy for transport predictions.

Cross-Spring Connections

  • → airSpring: f64 GPU dispatch batching pattern
  • → wetSpring: FusedMapReduceF64 pattern for bulk statistics; Anderson localization shared primitives
  • → ToadStool: 195 acceptance checks, 6 bugs found
  • → neuralSpring: isomorphic GEMM serves plasma and nuclear
  • → groundSpring: spectral primitives + QCD inverse problems

baseCamp Papers

Papers 07, 10, 15, 25 — see baseCamp Science for full list.