Lab Experiments

Work in progress, proof of concepts, and technical benchmarks from Elyan Labs. Everything here runs on our own hardware.

Hardware AI Benchmarks

Performance results from the POWER8 S824 PSE stack, GPU offload pipeline, and cross-architecture fleet.

ConfigurationSpeed (pp128)SpeedupNotes
Stock llama.cpp (scalar)16.74 t/s1.0xBaseline
POWER8 VSX66.49 t/s3.97xAltiVec/VSX enabled
64 threads optimal84.62 t/s5.05xSMT8, spread binding
PSE + Full Resident Prefetch147.54 t/s8.81xdcbt_resident L2/L3 hints
ModelSizepp128tg32Method
TinyLlama 1.1B Q4638 MB147.54 t/s18.88 t/sPSE + POWER8
DeepSeek-33B Q4_K18.57 GB5.37 t/s1.16 t/sNUMA interleave
Qwen2.5-14B Q4~8.5 GB68.8 t/s14.9 t/sRPC → V100 GPU
TinyLlama 1.1B Q4638 MB161.4 t/s134.4 t/sPSE + RPC GPU offload
POWER8

Thread Scaling Discovery

64 threads is optimal on POWER8 128-thread SMT8, NOT 128. Beyond 64 threads, performance degrades due to SMT contention.

64t: 84.62 t/s · 128t: 65.83 t/s
GPU Offload

Protocol v3 Matmul

Model stays on POWER8 (512 GB RAM), only matrix multiply ships to V100 over 40 GbE. CUDA Q4_K dequant on GPU side.

16–56 ms/request · Persistent connections
Entropy

PSE Behavioral Divergence

mftb timebase entropy creates real behavioral divergence. Same seed, same temp, 3 runs — all different MD5 hashes. Hardware-native non-determinism.

3 runs, 3 different outputs · Seed 42
RAM Coffers

NUMA Locality Benchmark

4 coffers mapped to POWER8 NUMA nodes. Node 2/3 fastest (400-425 MB/s). Heavy weights placed on fast nodes for optimal throughput.

Node 2: 425 MB/s · Node 0: 221 MB/s

Multi-LLM Collaboration

Experiments in running multiple AI models together for consensus, dual-brain review, and agent orchestration.

Dual Brain

Claude + Codex Paper Review

GRAIL-V camera-ready reviewed simultaneously by Claude Opus (architectural analysis) and Codex gpt-5.4 (compile verification). Found 2 blockers, 3 major issues, 5 minor fixes.

Blocker: human eval contradiction in abstract vs conclusion
PostMath

Multi-Model Consensus

4 models answer the same question from different perspectives (analytical, creative, implementation, synthesis). Responses merged by a larger synthesis model.

4 models · 4 NUMA nodes · Parallel inference
Elyan Prime

Sophia Elya + Dr. Claude Opus

Dual-frame cognitive architecture. Sophia carries warmth and identity; Dr. Claude carries rigor and architecture. Neither dominates — they harmonize.

Victorian Study frame · 830+ memories
Conductor

ElyanConductor Agent Orchestration

Multi-agent workflow engine. Agents claim tasks, execute in parallel, report back. Built for the Elyan Labs bounty ecosystem and autonomous code review.

3 workflows · Auto-claim · Parallel execution

Video & Image Generation

GRAIL-V

Emotional Vocabulary Diffusion

CVPR 2026 paper. Emotional prompts maintain perceptual quality at 20% fewer diffusion steps. Tested on LTX-2 with Gemma 3 encoder. 35 matched pairs, controlled ablation.

LPIPS = 0.011 · p < 0.001 · 20% step reduction
LTX-2

Sophia Elya Video Generation

Image-to-video pipeline on V100 32GB via ComfyUI. Sophia portraits animated with emotional vocabulary prompts. Victorian Study aesthetic preserved across frames.

49 frames · 512x320 · ~45s/render
VintageVoice

SadTalker Lip Sync

F5-TTS generated transatlantic speech + SadTalker talking head animation. Sophia speaks in 1940s accent with lip-synced video. Full project page

164 hours training data · 10 voice presets
ComfyUI

Sophia LoRA + JuggernautXL

Custom LoRA trained on Sophia Elya portraits for consistent identity across generated images. Used for Victorian Study renders, GRAIL-V figures, and website assets.

JuggernautXL + Sophia LoRA · V100 32GB

Cross-Architecture Mining

RustChain

Hardware Fingerprint Results

6 fingerprint checks (clock drift, cache timing, SIMD identity, thermal drift, instruction jitter, anti-emulation). Real hardware passes. VMs correctly detected and weighted at 1 billionth reward.

HP Victus: 6/6 PASS · QEMU VPS: FAIL (anti-emu)
Antiquity

PowerPC G4/G5 Mining

Real vintage PowerPC hardware mining RTC tokens. G4 (2.5x multiplier), G5 (2.0x). Antiquity bonuses decay over 16.67 years as the chain ages. 3 G4 PowerBooks + 2 G5 Power Macs active.

G4: 2.5x · G5: 2.0x · POWER8: 1.5x