Skip to content

README.md

Latest commit

 

History

History
121 lines (95 loc) · 6.63 KB

README.md

File metadata and controls

121 lines (95 loc) · 6.63 KB
Raw

baseCamp/systems — Novel NPU Systems and Multi-Tenancy

Date: February 27, 2026 Central question: How many concurrent systems can a single AKD1000 handle? Answer: 7 fully independent systems simultaneously, or 11+ outputs from one program.

The Hardware Reality

BrainChip markets the AKD1000 as a single-task inference accelerator. The silicon tells a different story.

AKD1000 physical budget:
  1,000 NPs total (80 DP-NPs + 920 general NPs)
  4 SRAM types: filter (64-bit), threshold (51-bit), event (32-bit), status (32-bit)
  SkipDMA: NP-to-NP without PCIe round-trip
  set_variable(): readout swap in 86 µs (not 14 ms — quantization-matched)
  Batch=8 sweet spot: 2.4× throughput from amortized PCIe
  Hardware determinism: same input → same output, always
  16 GB BAR1 address space: full NP mesh may be memory-mapped

The SDK never combines these into a multi-tenant view. That is the gap.

Simultaneous System Packing

Every model below has been validated on real AKD1000 or has a confirmed architecture that fits within NP budget constraints:

NP budget: 1,000 total
─────────────────────────────────────────────────────────────────────────────────────────────
 Slot  System                         NPs  NP Start  NP End  Status        Throughput
─────────────────────────────────────────────────────────────────────────────────────────────
  1    ESN QCD thermalization         179  0x0000    0x00B3  ✅ validated    18,500 Hz
  2    Transport predictor (D*,η*,λ*) 134  0x00B3    0x0139  ✅ validated    17,800 Hz
  3    DS-CNN keyword spotting        220  0x0139    0x0215  📋 analysis     ~1,400 Hz
  4    ECG anomaly detection           96  0x0215    0x0275  📋 analysis     ~2,200 Hz
  5    Phase classifier (SU3)          67  0x0275    0x02B8  ✅ validated    21,200 Hz
  6    Anderson regime classifier      68  0x02B8    0x02FC  ✅ validated    22,400 Hz
  7    Minimal sentinel (50-dim)       50  0x02FC    0x032E  ✅ confirmed    ~24,000 Hz
─────────────────────────────────────────────────────────────────────────────────────────────
  TOTAL                               814                    186 NPs spare
─────────────────────────────────────────────────────────────────────────────────────────────

NP addresses are cumulative (end of slot N = start of slot N+1). Validated by cargo run --bin bench_exp002_tenancy — no overlaps, 186 spare NPs confirmed.

7 distinct systems on one chip, simultaneously. The 186 spare NPs fit an 8th minimal system or a temporal integration ESN.

With the 11-head pattern (SkipDMA, single program):

1 reservoir program (179 NPs) → 11 independent output heads
  Head 0: thermalization flag
  Head 1: phase classification
  Head 2: anomaly score
  Head 3: β priority (pre-scan)
  Head 4: CG iteration estimate
  Head 5: rejection likelihood
  Head 6: quality score
  Head 7: next-run recommendation
  Head 8: deconfinement order parameter
  Head 9: transport coefficient D*
  Head 10: transport coefficient η*

→ 11 outputs at single-program latency (54 µs, not 11 × 54 µs)

Novel Systems (Not in Any SDK Zoo)

System NPs Capability Status
Multi-tenancy controller 814 7 simultaneous domains scaffolded
Online evolution engine 179+50 136 gen/sec live adaptation scaffolded, ✅ validated (wetSpring)
NPU conductor (11-head) 179 11 physics outputs, 1 program ✅ validated (hotSpring Exp 023)
Chaotic attractor tracker 259 ESN for Lorenz/Rössler/MSLP Architecture + NeuroBench ref
Temporal PUF 68 Hardware fingerprinting, 6.34 bits scaffolded, ✅ validated (wetSpring NPU)
Adaptive edge sentinel 179+50 Domain-shift detection + adapt scaffolded, ✅ validated (wetSpring)
Neuromorphic PDE solver 200-400 Poisson/heat via multi-pass FC Planned
Physics surrogate ensemble 560 4-domain co-located surrogates Architecture defined

Hybrid executor: HybridEsn is complete; SRAM weight verification via verify_load().

Extended SDK Use Cases

BrainChip claims rustChip demonstrates
AkidaNet: 65% ImageNet top-1 Domain adaptation at 86 µs, online fine-tune, 1000→N class hot-swap
DS-CNN: 93.8% keyword spotting Multi-vocabulary hot-swap, simultaneous KWS+anomaly, acoustic sentinels
DVS: 97.9% gesture Continuous temporal tracking via ESN, multi-gesture ensemble
ECG: 97.4% anomaly 12B inferences on coin cell, real-time HAB/environmental adaptation
YOLO: mAP 0.42 Temporal object persistence via ESN integration, multi-class co-location

Reading Order

For hardware validation (you have a live AKD1000):

  1. multi_tenancy.md — can the chip really hold all 7 simultaneously?
  2. bench_multi_tenancy binary — measures it directly
  3. npu_conductor.md — run the 11-head system

For novel capabilities:

  1. online_evolution.md — the most surprising capability
  2. temporal_puf.md — hardware fingerprinting with no extra software
  3. chaotic_attractor.md — the attractor tracking / MSLP NeuroBench bridge

For the outreach case:

  1. The "Beyond SDK" extended analyses in models/edge/beyond_sdk/
  2. physics_surrogate.md — the clearest GPU+NPU value demonstration
  3. metalForge/experiments/002_MULTI_TENANCY.md for empirical validation