Silver Layer¶
Derived tables built from the bronze stack. The silver layer transforms raw spots into embeddings and aggregated signatures for downstream analysis and model training.
Prerequisites¶
- Bronze stack fully populated (see Bronze Stack)
solar.bronzepopulated (required for all JOINs)
Step 1: Generate CUDA Embeddings¶
The bulk-processor generates float4 embeddings from WSPR spots joined with
solar indices, stored in wspr.silver.
Requires NVIDIA GPU
The bulk-processor requires an NVIDIA GPU with sufficient VRAM. The RTX PRO 6000 (96 GB) processes all 10.8B spots in a single pass.
Not in RPM
bulk-processor is not yet packaged in the ionis-cuda RPM.
Build locally: cd ionis-cuda && mkdir build && cd build && cmake .. && make
Verification:
The v_quality_distribution materialized view auto-populates as rows are
inserted into wspr.silver:
Step 2: Build Aggregated Signatures¶
Signatures compress 10.8B raw spots into ~93M median-bucketed entries — a 115:1 compression ratio that strips site-level noise and reveals the atmospheric transfer function.
See Aggregated Signatures for full methodology and per-band distribution.
bash /usr/share/ionis-core/scripts/populate_signatures.sh
# Or with custom host:
# CH_HOST=10.60.1.1 bash /usr/share/ionis-core/scripts/populate_signatures.sh
Verification:
clickhouse-client --query "SELECT count() FROM wspr.signatures_v2_terrestrial"
# Expected: ~93,600,000
QA Actuals¶
Clean-slate rebuild on 9975WX (2026-02-07):
Table Rows Time
---------------------- -------------- ---------
wspr.silver 4,430,000,000 ~45 min
wspr.signatures_v2_terrestrial 93,600,000 3m31s
v_quality_distribution ~6,100,000 (auto)
Next Steps¶
- Gold layer: See Gold Layer for training tables and CSV export
- Training: See Training for model architecture and training methodology