RX 580 Vulkan hits 16 t/s ceiling on llama.cpp

The RX 580 Vulkan experiment exposes a real gap: theoretical memory bandwidth (256 GB/s) vs. actual utilization (~15%), revealing how critical hardware matrix ops are for LLM inference throughput.

• –Polaris (gfx803) has no fp16, bf16, or int dot product acceleration in RADV — every matrix multiply runs as a generic fp32 compute shader, which is massively inefficient for transformer attention patterns
• –The gap between theoretical ~100 t/s (bandwidth-bound) and actual ~16 t/s is the real cost of missing tensor core equivalents on older AMD hardware
• –ROCm with HIP (DGGML_HIPBLAS=ON targeting gfx803) is the realistic path forward — Vulkan lacks the low-level primitives to close this gap on Polaris
• –llama.cpp's Vulkan backend is solid for supported hardware but cannot compensate for missing ISA features; no amount of flag tuning helps
• –This is a useful data point for anyone evaluating old AMD GPUs for local inference — Vulkan is not a universal fallback that extracts full hardware performance