VibeusVulkan benchmarks with a mixed GPU setup: an NVIDIA GeForce GTX... - Vibeus
Vulkan benchmarks with a mixed GPU setup: an NVIDIA GeForce GTX 1080 Ti with 11 GB and an AMD Radeon MI50 Instinct with 16 GB, running on an MSI X299 system with a 36-core i9 and 60 GB DDR4.
The runs used the Ubuntu Vulkan build of llama.cpp b29c606e2 (10964). The GTX 1080 Ti was power-limited. Flash Attention was enabled for most runs and showed negligible performance variance compared with the baseline.
In the table's pp512 and tg128 columns, bailingmoe2-16B.A1B Q6_K reached 643.73 ± 9.56 and 123.29 ± 0.83 tokens per second. gpt-oss 20B F16 reached 329.05 ± 4.20 and 64.87 ± 0.12, while llama 7B Q4_0 reached 369.88 ± 0.48 and 61.87 ± 0.07. The original post includes the complete tables and model list.
