Skip to content

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.bronze populated (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.

bulk-processor --host 192.168.1.90

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:

clickhouse-client --query "SELECT count() FROM wspr.silver"
# Expected: ~4,430,000,000

The v_quality_distribution materialized view auto-populates as rows are inserted into wspr.silver:

clickhouse-client --query "SELECT count() FROM wspr.v_quality_distribution"
# Expected: ~6,100,000

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