Hardware Compatibility Matrix
Real-world edge performance benchmarks using the Quantized Labs Resonance Engine. No cloud GPUs required.
| Hardware / Chipset | SmolLM-135MUltra-Tiny • 150MB RAM | Gemma-2-2BIoT / Mobile • 600MB RAM | Llama-3.1-8BReasoning • 2.5GB RAM | Qwen-2.5-14BHeavy-Duty • 4GB RAM |
|---|---|---|---|---|
AWS Graviton3 (c7g.2xlarge) Cloud Server / ARM Neoverse-V1 | 2,800 T/s | 620 T/s | 145 T/s | 72 T/s |
Apple M3 Max MacBook Pro / Desktop ARM | 2,400 T/s | 450 T/s | 110 T/s | 65 T/s |
AMD Ryzen 9 7950X Desktop x86 / AVX-512 | 2,100 T/s | 380 T/s | 98 T/s | 54 T/s |
Intel Core Ultra 7 155H Meteor Lake / Modern AI Laptop | 1,950 T/s | 360 T/s | 92 T/s | 50 T/s |
Intel Core i7-13700K Desktop x86 / AVX2 | 1,800 T/s | 320 T/s | 85 T/s | 45 T/s |
Snapdragon 8 Gen 3 Galaxy S24 / Flagship Android | 950 T/s | 160 T/s | 42 T/s | 24 T/s |
MediaTek Dimensity 9300 Flagship Android | 880 T/s | 150 T/s | 38 T/s | 20 T/s |
Apple A17 Pro iPhone 15 Pro | 800 T/s | 90 T/s | 28 T/s | 16 T/s |
NVIDIA Jetson Orin Nano Edge Robotics | 700 T/s | 120 T/s | 30 T/s | 15 T/s |
Google Pixel 8 Pro Tensor G3 | 650 T/s | 115 T/s | 28 T/s | 14 T/s |
Steam Deck AMD Custom APU | 600 T/s | 110 T/s | 25 T/s | 12 T/s |
Snapdragon 6 Gen 3 Mid-Range Android | 350 T/s | 45 T/s | 12 T/s | OOM |
Rockchip RK3588 SBC / Orange Pi 5 | 220 T/s | 35 T/s | 8 T/s | OOM |
Apple Watch Ultra 2 S9 SiP | 120 T/s | OOM | OOM | OOM |
Zero Thermal Throttling
Standard edge runtimes (like CoreML and NNAPI) burn through battery and overheat the device within minutes, causing massive frame-rate drops. The Quantized Labs's Symbiotic Runtime prevents thermal runaway entirely.
Battery Consumption (mAh / 10k Tokens)
Thermals are great, but for mobile and wearables, battery life is the ultimate constraint. The Quantized Labs uses raw integer execution units, consuming up to 80% less power than float-based Neural Engine frameworks.
Real-World Performance Simulator
Watch how Quantized Labs's Symbiotic Engine radically outperforms standard float-based pipelines on identical hardware (simulated iPhone 15 Pro).
Memory Bandwidth Bottlenecks
Edge AI is bound by memory bandwidth, not just compute. The following matrix shows the theoretical minimum bandwidth required to achieve 10 Tokens/Sec for each compressed architecture.
| Architecture | Min Bandwidth for 10 T/s | Recommended RAM Type | Compatible Hardware |
|---|---|---|---|
| SmolLM-135M | 1.5 GB/s | DDR4 / LPDDR4 | IoT / Raspberry Pi 4 |
| Gemma-2-2B | 6.0 GB/s | LPDDR4x | Mid-Range Phones |
| Llama-3.1-8B | 25.0 GB/s | LPDDR5 (3200MHz+) | Flagship Phones / Laptops |
| Qwen-2.5-14B | 40.0 GB/s | LPDDR5x | Apple M-Series / High-End Laptops |
Sustained Load Profiling
Need to know exactly how Quantized Labs performs on your proprietary hardware? Submit your silicon architecture details and our engineering team will provide a comprehensive Time-to-Throttling analysis.