Real-World Performance — 2026-02-16
Absolute wall-clock estimates for workloads on Triton VM.
Proving Cost Model
Padded height = next_power_of_2(max table height across 6 AETs).
Formula: padded_height * 300 * log2(ph) * 3ns (optimistic lower bound).
Real-world: 2-5x formula. GPU: 3-10x speedup over CPU.
| Padded Height | CPU Estimate | RAM |
|---|---|---|
| 2^16 | 1-3s | ~6 GB |
| 2^20 | 30-90s | ~100 GB |
| 2^22 | 3-8 min | ~800 GB |
| 2^24 | 15-40 min | ~3 TB+ |
Feasibility Summary
| Workload | Cycles | Padded | GPU Time | Feasible |
|---|---|---|---|---|
| Hash/Merkle proof | <100K | 2^17 | <1s | Yes |
| Token transfer | <500K | 2^19 | <5s | Yes |
| Quantum 2-5 qubits | <20K | 2^14 | <1s | Yes |
| RLWE ciphertext add | <50K | 2^15 | <1s | Yes |
| MNIST MLP (784-128-10) | ~2.2M | 2^22 | 30-90s | GPU only |
| Quantum 10-12 qubits | ~500K | 2^19 | 5-15s | Yes |
| Quantum 14-16 qubits | ~30M | 2^25 | 30-90 min | Marginal |
| FHE multiply N=1024 (opt NTT) | ~1.3M | 2^21 | 10-30s | GPU only |
| FHE multiply N=1024 (current) | ~567M | 2^30 | Hours | No |
| Transformer attention head | ~12M | 2^24 | 2-6 min | Marginal |
Key Bottleneck
poly.tri NTT wastes ~500x cycles from bounded loop overhead. Fix: restructure to direct butterfly addressing. This alone makes small FHE marginally feasible.
Competitive Position
- 100-200x behind SOTA for AI inference (zkPyTorch, EZKL)
- Sweet spot: hash-heavy crypto, small provable computations, recursive proofs
- Novel niche: verifiable quantum simulation (2-12 qubits)
- Dual-target advantage (CUDA + zkVM from same source) is unique