Sovereign Fleet Mathematical Topography
Empirical Kernels · KV Memory Manifolds · Vector Singularities · Sunk Economics
OMEN: Dual Arc B70 (vLLM XPU, FlashAttn, 782 tok/s)
AM4: RTX 5070 + 4070 Ti TP2 (2,920 tok/s)
FX99: Coder-7B Utility (Sub-second Guardian)
1. Prefill Latency Topography: T_prefill(L)
Native XPU Flash Attention O(L) vs. Resident Triton Attention O(L²) on Intel Arc B70
Triton Attention (Resident)
Native Flash Attention (Promoted)
Range: 0 to 45,000 Prompt Tokens
T_prefill, Flash(L) ≈ 0.00067 · L | T_prefill, Triton(L) ≈ 0.0035 · L + 0.0000035 · L² | 39K Depth: 26.0s vs. 661.0s (25.4× Gain)
2. KV Cache Concurrency: N_seq(L)
Memory Utilization Fraction μ vs. Context Length
VRAM Util (μ):
Pool(0.90) = 95,783 tokens (133 blk) | Short (4K): N=23.9 seqs | Deep (38K): N=1.77 seqs
3. Speculative MTP Depth: R_eff(k)
Throughput Scaling vs. Verification Overhead
k=2 Sweet Spot: 56.5 tok/s (+17% gain) | p1=0.97, p2=0.85 | k=3 verification penalty drops rate
4. JEPA Latent Representation Space
Collapsed 1D Ray (OT002) vs. Spherical Spread (InfoNCE)
Collapse: cos(z_i, z_j) → 1.0000 (127/128 dead) | InfoNCE: 0/128 collapsed, 44.4% R@3
5. The Two Economies Cost Surface
Sunk Local Electricity vs. Metered Cloud Margin
Monthly Volume:
Local Fleet: ~$42.12/mo flat (450W @ $0.13/kWh) | Cloud API: $750.00/mo ($15/M tok)
6. Sovereign Fleet Physical Boundary Matrix & Tripwires
Empirically Mapped Hard Failure Ceilings across OMEN, AM4, and FX99
| Subsystem / Node | Operational Parameter | Safe Envelope | Critical Tripwire | Observed Failure Mode | Codified Mitigation / Law |
|---|---|---|---|---|---|
| OMEN Dual B70 | VRAM Junction Temperature | < 80°C | ≥ 95°C | Thermal throttling, driver lockup, dropped PCI commands | Hard tripwire abort; resume strictly locked until < 80°C |
| OMEN Dual B70 | Driver Idle Decay (ADR-0043) | < 20s between calls | > 60s idle (Shallow) / >15m (Deep) | Power-gating drops decode rate 4×; deep collapse drops to 33% | Issue 1-token warmup ping every 20s; deep collapse requires full restart |
| OMEN Dual B70 | Multi-Tenancy (ADR-0041) | N = 1 tenant / card | N ≥ 2 co-resident models | Incumbent decode rate plunges from 105 tok/s to 28.39 tok/s permanently | Co-residency poisons incumbent: isolate seats across cards |
| OMEN Dual B70 | PCIe Device Enumeration | Immutable BDF addresses | Device Index 2 Naive Alloc | DirectX/Vulkan index maps to Intel iGPU; discrete card sits idle | Select devices strictly by immutable PCIe BDF address, never by index (ADR-0042) |
| AM4 Host Node | Root NVMe Disk Utilization | < 90% (50 GiB free) | ≥ 97% (18 GiB free) | Filesystem read-only lock; failed docker mounts; service termination | Zero weight copies to AM4; symlink across shared partitions strictly |
| FX99 Worker Node | Tool-Use VRAM Envelope | Summaries & Formatting | Complex Multi-Tool Calling | 8GB VRAM 7B model truncates JSON schema; citation bleed | Isolate FX99 to Stage 3 Projection & Schema Guardian; zero tool calls |
| Batch Scheduler | VRAM Thrash / Swaps (CP-SAT) | 3 swaps (0 misses) | 24 swaps (19 misses, FIFO) | 4+ minutes wasted in PCIe weight transfer per 250-job batch | CP-SAT constraint optimization achieves optimal makespan (87.5% reduction) |