CMP 100-210 powers Medina 14B inference

The title oversells parity with the 3090, but the underlying result is still useful: this is a cheap, awkward mining card that becomes surprisingly practical when you care more about VRAM and adapter switching than peak tok/s.

• –The CMP 100-210 posts 47.3 tok/s versus 78.1 tok/s on the RTX 3090, so it is not a throughput equal.
• –LoRA hot-swap latency stays in the 13-18 ms range across all tested GPUs, which makes the workflow feel CPU/I/O-bound more than GPU-bound.
• –The card’s 16 GB VRAM is enough for a 36K context setup, but the 3090 still buys real safety margin and faster inference.
• –Multi-GPU layer splitting underperforms the solo 3090 because PCIe synchronization overhead eats the benefit.
• –For local LLM builders on a budget, the CMP 100-210 is interesting as a dedicated inference node, not as a universal 3090 replacement.