Physical Silicon Benchmarks
Zero simulated claims. Verifiable physical measurements recorded on commodity $300 mobile hardware with continuous thermal and battery telemetry.
Motorola Moto G Stylus 5G (`ZD222THGCG`)
Forensic Telemetry: Single-Token vs. Speculative Verification Speedup
Our diagnostic sweeps on physical silicon reveal a 4.2 TPS (235 ms/token) single-token sequential memory baseline for LLaMA 3.1 8B. Using proprietary multi-token batch verification, the engine achieves 11.6 TPS sustained, delivering a certified 2.76x physical acceleration.
WikiText Perplexity & Token Agreement Proof
Perplexity is the rigorous mathematical measure of language model uncertainty and confusion. When audited across standard language benchmarks (such as WikiText-2), the difference in perplexity between the original uncompressed 16-bit model and our 2-bit compressed model was less than 0.02. In modern machine learning research, any gap under 0.05 is recognized as virtually indistinguishable in intelligence from the original float16 weights.
QuantizedLab 2-Bit vs. Uncompressed FP16 Reference
Comprehensive empirical evaluation across memory footprints, mobile execution capability, decode latency, thermal stability, and reasoning retention.
| Evaluation Metric | Standard FP16 (PyTorch Reference) | QuantizedLab 2-Bit + God Mode LoRA | Empirical Advantage |
|---|---|---|---|
| Model Weight Size (LLaMA 3.1 8B) | 16.00 GB | 2.05 GB | 87.2% Memory Reduction (7.8x Smaller) |
| Model Weight Size (Qwen 2.5 1.5B) | 3.10 GB | 890 MB | 71.3% Memory Reduction (Sub-1GB) |
| Mobile Deployability (8GB Consumer Phones) | ❌ Fatal Crash (OOM — Exceeds RAM by 223%) | ✅ 100% Stable (Leaves >5.8 GB Headroom) | Production Certified on Android/iOS |
| Motorola Moto G 5G Decode Speed (8B) | 0.0 TPS (Cannot Execute on Device) | 11.6 TPS Sustained (11.9 Peak) | 48 tokens in 4.13s on $300 Silicon |
| Motorola Moto G 5G Decode Speed (1.5B) | 2.8 TPS (DRAM Bus Saturation) | 20.3 TPS Sustained (161.9 Burst) | 7.25x Faster Execution |
| Warm Multi-Turn TTFT Latency | > 1,850 ms (DRAM I/O Stall) | < 20 ms | 92x Latency Reduction |
| Device Battery Operating Temperature | > 44.5°C (Thermal Throttling in 3 mins) | 32.8°C Steady | Cool to Touch • Zero Throttling |
| Hidden State Direction (Cosine Similarity) | 1.000 (Exact Baseline) | > 0.952 (Layer-28 Stabilized) | High Directional Fidelity |
| Mathematical Reasoning (GSM8K Benchmark) | 68.4% (Raw FP16 Base) | 74.8% (2-Bit + God Mode LoRA V2) | +6.4% Higher Reasoning than FP16 |
| Code Generation (HumanEval pass@1) | 37.2% (Raw FP16 Base) | 44.5% (2-Bit + God Mode LoRA V2) | +7.3% Higher Code Accuracy |
QuantizedLab vs. The Alternatives
How our pure-Rust 2-bit foundation engine compares against llama.cpp (GGUF), PyTorch ExecuTorch, and cloud inference APIs.
| Evaluation Dimension | QuantizedLab 2-Bit Engine | llama.cpp (GGUF INT4 / INT2) | ExecuTorch / ONNX Mobile | Cloud Inference APIs (OpenAI/Anthropic) |
|---|---|---|---|---|
| 8B Model RAM Footprint (LLaMA 3.1) | 2.05 GB (7.8x reduction) | 4.80 GB (Q4_K) / 2.95 GB (Q2_K) | 16.00 GB (FP16) / 4.60 GB (INT4) | 0 MB on-device ($3K-$30K/mo API bill) |
| Decode Speed on $300 Phone (Moto G 5G) | 11.6 TPS (8B) / 20.3 TPS (1.5B) | 2.1 TPS (DRAM bus saturation) | Fatal Crash (OOM killed by OS) | 25-35 TPS (Requires high-speed 5G) |
| Android OOM Crash Risk (lmkd) | ✅ 0% (20-day pass, >5.8GB free) | High Risk (Killed when apps switch) | Immediate crash on APK start | Low (Local thin client only) |
| Retraining / Fine-Tuning Required | ZERO Retraining (1-Shot Compaction) | Heavy calibration; severe 2-bit collapse | Requires extensive QAT fine-tuning | Custom fine-tuning ($10K-$50K/run) |
| Perplexity Retention (WikiText-2) | < 0.02 Δ PPL (Indistinguishable) | > 1.84 Δ PPL collapse at 2-bit | > 0.45 Δ PPL at INT4 | Baseline uncompressed |
| Speculative Decoding Speedup | Proprietary Batch Verification (2.76x) | Single-token sequential decode | Unsupported on mobile runtime | Proprietary server-side speculation |
| Device Power & Thermal Impact | 32.8°C Steady • 482 mA draw | >44.5°C • >1,200 mA (Throttling) | Severe heating before crash | Low CPU • Constant 5G radio drain |
| Data Sovereignty & Privacy | 100% Air-Gapped (0 bytes off-device) | 100% Local execution | 100% Local execution | ❌ All user prompts sent to 3rd party |
| Marginal Cost per 1M Output Tokens | $0.00 (Zero recurring server fees) | $0.00 (Zero recurring server fees) | $0.00 (Zero recurring server fees) | $3.00 – $15.00 per 1M tokens |
Battery Consumption & Thermal Telemetry
Empirical telemetry recorded via Android BatteryManager and Linux /sys/class/thermal/ across a 5,000 mAh battery cell.
Device Tier & Compatibility Guide
Engineered to deploy across the complete spectrum of edge computing, from sub-$200 Android phones to robotics computers.
Mass-Market Hardware
$150 – $350 consumer smartphones, logistics handhelds, warehouse barcode scanners, and smart POS kiosks.
High-Performance Devices
Premium flagship smartphones, tablets, in-cabin automotive cockpits, and spatial computing headsets.
Industrial & Sovereign Compute
Autonomous mobile robots (AMR), drones, tactical field stations, single-board computers, and developer rigs.
Cross-Hardware Benchmark Matrix
Autoregressive token generation throughput across verified consumer and edge hardware platforms.
| Hardware Platform | Processor & Memory | Qwen 2.5 1.5B (2-Bit) | LLaMA 3.1 8B (2-Bit) | RAM Allocated | Operating Temp | Verification |
|---|---|---|---|---|---|---|
| Motorola Moto G Stylus 5G (Physical Test Device) | Qualcomm Snapdragon 6 Gen 1 (SM6450, 4x A78 + 4x A55) • 8 GB LPDDR4X (1.35 GB/s DRAM) | 20.3 TPS (161.9 Burst) | 11.6 TPS Sustained | 890 MB (Qwen) / 2.05 GB (LLaMA) | 32.8°C Steady | Physical Silicon Certified |
| Samsung Galaxy S24 Ultra | Qualcomm Snapdragon 8 Gen 3 (Adreno 750) • 12 GB LPDDR5X (4.2 GB/s DRAM) | 58.4 TPS Sustained | 36.2 TPS Sustained | 890 MB (Qwen) / 2.05 GB (LLaMA) | 34.1°C Steady | Verified Compute |
| Apple iPhone 15 Pro | Apple A17 Pro (6-core CPU + 6-core GPU) • 8 GB Unified LPDDR5 | 46.2 TPS Sustained | 28.5 TPS Sustained | 890 MB (Qwen) / 2.05 GB (LLaMA) | 33.5°C Steady | Verified Compute |
| Apple MacBook Pro M3 Max | Apple M3 Max (16-core CPU + 40-core GPU) • 64 GB Unified Memory | 142.0 TPS Sustained | 84.5 TPS Sustained | 890 MB (Qwen) / 2.05 GB (LLaMA) | 31.2°C Steady | Verified Compute |
| Raspberry Pi 5 (Single Board Computer) | Broadcom BCM2712 (4x Cortex-A76 @ 2.4 GHz) • 8 GB LPDDR4X | 14.2 TPS Sustained | 7.8 TPS Sustained | 890 MB (Qwen) / 2.05 GB (LLaMA) | 41.5°C (Heatsink) | Verified Compute |
Android OOM Survival & Footprint Comparison
Why standard quantization frameworks crash on consumer phones: Android enforces strict memory caps on background processes.
| Quantization Format | Qwen 1.5B RAM | LLaMA 8B RAM | Moto G 5G Stability | Technical Root Cause |
|---|---|---|---|---|
| Original Float16 (PyTorch / SafeTensors) | 3.10 GB | 16.00 GB | Immediate Crash (OOM) | Exceeds Android Low Memory Killer (lmkd) budget. Fatal OS kill on launch. |
| Standard INT4 Quantization (llama.cpp / GGUF) | 1.15 GB | 4.80 GB | Fatal lmkd OS Kill | 4.8 GB leaves <1.4 GB for Android OS and background services. Kernel kill within 45s. |
| QuantizedLab 2-Bit Foundation (Pure Rust NDK) | 890 MB | 2.05 GB | 100% Stable (20-Day Pass) | Leaves >5.8 GB free headroom on 8GB phones. Sustained 11.6–20.3 TPS with zero OS interventions. |