From GTX-1080Ti to dual GPU with Radeon MI50

The Nvidia GTX-1080Ti has 11GB of VRAM and capable for running 8B to 14B local models. I added an AMD Radeon MI50 16GB VRAM. Now with 27GB VRAM I'm able to run most 30B to 35B local models from VRAM. I got the MI50 GPU for less than $140 and it's firmware flashed to Radeon VII so mini Display port work but power levels are lowered. Benchmarked using Llama.cpp Ubuntu Vulkan pre-built binary build: 851cb34f2 (11055) No driver hassles or conflicts thanks to Kubuntu and Vulkan. preview.redd.it/efoht9forjqh1.png I ran GTX-1080Ti by itself which causes CPU/GPU offloading (GGML_VK_VISIBLE_DEVICES=1) . Then ran same LLM models with dual GPU configuration. I added command sleep 10 to give system time to get ready for next run. Here is the command I used: GGML_VK_VISIBLE_DEVICES=1 time ~/llama-b11055/llama-bench -fa on -ngl 99 -m /gemma-4-26B-A4B-it-UD-Q6_K_XL.gguf && sleep 10 && time ~/llama-b11055/llama-bench -fa on -ngl 99 -m /gemma-4-26B-A4B-it-UD-Q6_K_XL.gguf preview.redd.it/9bs3qhffrjqh1.png GGUF Models and Sizes Qwen3.6-35B-A3B-MXFP4_MOE.gguf (20.65 GiB) NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Q5_K_M.gguf (25.10 GiB) Qwen3.8-27B-Q6_K.gguf (21.30 GiB) medgemma-27b-it-UD-Q6_K_XL.gguf (22.09 GiB) gemma-4-26B-A4B-it-UD-Q6_K_XL.gguf (21.68 GiB) Single GPU Results (Visible Devices = 1 GTX-1080Ti) model size params pp512 tg128 qwen35moe 35B.A3B MXFP4 MoE 20.65 GiB 35.51 B 248.10 12.60 nemotron_h_moe 31B.A3.5B Q5_K - Medium 25.10 GiB 32.91 B 230.64 10.78 qwen35 27B Q6_K 21.30 GiB 27.32 B 56.10 2.12 gemma3 27B Q6_K 22.09 GiB 27.01 B 62.73 1.51 gemma4 26B.A4B Q6_K 21.68 GiB 25.23 B 162.12 11.76 Dual GPU Results (Faster GTX-1080Ti and Radeon MI50) model size params pp512 tg128 qwen35moe 35B.A3B MXFP4 MoE 20.65 GiB 35.51 B 333.79 31.78 nemotron_h_moe 31B.A3.5B Q5_K - Medium 25.10 GiB 32.91 B 298.69 65.02 qwen35 27B Q6_K 21.30 GiB 27.32 B 124.85 15.60 gemma3 27B Q6_K 22.09 GiB 27.01 B 127.58 14.96 gemma4 26B.A4B Q6_K 21.68 GiB 25.23 B 340.86 52.03 Dense models benefited the most: Gemma3 27B saw a staggering ~896% increase in generation speed (from 1.5 t/s to 14.96 t/s). Qwen35 27B saw a ~642% increase (from 2.1 t/s to 15.6 t/s). Nemotron 31B saw a ~504% increase . Great results considering cost less than $140 to add a 16GB Vram GPU. Now running 30B size Dense models is a really option.