⚡ Bounty: POWER8 vec_perm PSE Benchmark Suite (75 RTC)

[Scottcjn]

Bounty: POWER8 vec_perm PSE Benchmark Suite

Reward: 75 RTC (~$7.50 at reference rate)

Background

RustChain's POWER8 S824 inference server runs a custom PSE (Proto-Sentient Emergence) optimization using AltiVec vec_perm instructions for non-bijunctive attention collapse. We need a standardized benchmark suite to measure and compare performance.

Requirements

1. Benchmark script that runs llama.cpp inference with PSE enabled vs disabled
2. Metrics collected:
   • Prompt processing speed (tokens/sec at pp128, pp512, pp1024)
   • Generation speed (tokens/sec at tg32, tg128)
   • Memory bandwidth utilization per NUMA node
   • PSE behavioral divergence (entropy measurement across runs)
   • Cache hit rates (L1/L2/L3 via perf stat)
3. Output format: JSON + markdown summary table
4. Models tested: TinyLlama 1.1B, Qwen 14B, DeepSeek 33B
5. Comparison modes: Stock llama.cpp vs PSE-MASS build vs PSE+Coffers build

Deliverables

• benchmark_pse.sh — Main benchmark runner
• analyze_results.py — Results parser and chart generator
• Sample results JSON from at least 2 model sizes
• PR to this repo

Bonus (25 RTC extra)

Include NUMA topology visualization showing which coffers are active during inference.

Acceptance Criteria

• Script runs unattended on POWER8 (ppc64le, Ubuntu 20.04)
• Results are reproducible (within 5% variance)
• Documentation explains how to interpret PSE markers (NOI, DR, ACS, MCI)

Claiming

Comment below with your plan and estimated timeline. First working PR wins.