GMKtec EVO-X2 runs Llama 70B at 5 t/s

The GMKtec EVO-X2 marks a shift in local AI hardware, proving that high-bandwidth unified memory can finally replace expensive dual-GPU setups for 70B-class models.

• –5.25 t/s is a 3-4x performance leap over previous generation iGPUs (RDNA 3) which typically hit ~1.5 t/s on 70B models
• –128GB of LPDDR5X-8000 is the critical enabler, allowing 70B models to reside entirely in VRAM with room for large KV caches
• –Performance degradation at high context (dropping to 2.5 t/s at 16k) reveals the compute-to-bandwidth bottleneck inherent in current integrated solutions
• –Linux with ROCm 6.4.4 remains the optimal software stack for maximizing Strix Halo's gfx1151 architecture
• –The necessity of GRUB tweaks (ttm.pages_limit) highlights that "AI PCs" still require significant manual tuning for enthusiast-level local inference