GPU Compute — Live Evidence

Measured GPU compute performance on commodity hardware. DF64 precision, matmul throughput, vendor-agnostic via WebGPU/WGSL.

Validated on live hardware — strandGate RTX 3090, westGate RTX 4070. All numbers are measured, not theoretical.

Measured Performance (strandGate — RTX 3090)

BenchmarkResultMethod
DF64 matmul2,130 ops/sec512×512, measured via barraCuda.matmul
DF64 precision~14 significant digitsDouble-float emulation on FP32 ALUs
barraCuda capabilities98 methods LIVEGPU compute, linear algebra, FFT, SVD, ML
Shader languageWGSL (WebGPU)No CUDA dependency. Runs on any Vulkan GPU

What DF64 Is

DF64 (Double-Float 64) emulates f64 precision using pairs of f32 values. Consumer GPUs throttle native f64 to 1/32 or 1/64 of their FP32 rate. DF64 bypasses this by running two FP32 operations per logical f64 operation, achieving ~14 significant digits of precision at FP32 throughput rates.

Trade-off: DF64 uses 2x the FP32 ALU bandwidth per operation. The precision is real (verified against f64 reference). The throughput is lower than native f64 on datacenter GPUs (A100, H100) that have full-rate FP64 units. The advantage is running on $500 consumer hardware instead of $15,000 datacenter cards.

Lattice QCD — Multi-Vendor GPU vs CPU (strandGate)

SU(2) HMC (Hybrid Monte Carlo) lattice gauge theory. Same algorithm, same machine, both GPUs running identical WGSL shaders, cpu_mom validated path:

LatticeVolumeRTX 3090 msRX 6950 XT msCPU msBest Speedup
4^425617.27.4185.025.1x
8^44,09662.915.62,965.8190.0x

Omelyan 2MN integrator, n_md=20, dt=0.02. cpu_mom path (CPU-generated momenta, GPU molecular dynamics) after root-causing GPU PRNG polyfill bias.

Cross-GPU agreement: Both GPUs produce identical plaquette values within DF64 accumulated precision (|Δ|_GPU-GPU = 3.1×10⁻⁹ at 8^4 — five orders of magnitude below statistical error). Vendor-agnostic proof: same WGSL shaders, different silicon, identical physics.

Plaquette validation: |Δ|/σ < 1 vs CPU f64 reference at both lattice volumes. GPU molecular dynamics produces statistically identical physics to the CPU implementation.

Download the full trajectory data + provenance chain: hotSpring QCD pseudoSpore

GPU Workloads Running in Production

WorkloadStatusGate
Lattice QCD (HMC trajectories)LIVE — measured abovestrandGate
Matrix multiply (dense)LIVEstrandGate, westGate
SVD decompositionLIVEstrandGate
FFT (1D, 2D)LIVEstrandGate
AlphaFold MSA scoringCapacity assessedstrandGate
Neuromorphic (Akida)VFIO passthroughwestGate

Vendor Independence

barraCuda compute runs on any GPU with Vulkan 1.2+ support:

  • NVIDIA (tested: RTX 3090, RTX 4070)
  • AMD (tested: consumer Radeon via Vulkan)
  • Intel Arc (supported via Vulkan, not yet fleet-tested)

No CUDA. No ROCm. No vendor SDK. Pure WGSL shaders dispatched through the WebGPU API.

Pending: Live Benchmarks

This page currently shows static measurements. When petalTongue G19 Node Atomics rendering is complete, it will serve real-time benchmark results from barraCuda via spore-validate pt-render.

Data source: spore-validate nucleus strandGate --probe