2x AMD Radeon RX 7900 XTX

GPU

16

Run records

1

Models covered

4

Frameworks covered

L2 Cross-framework verified × 16

Evidence mix

Typical measured performance

Published benchmark averages per model × framework × quantization on this hardware.

  1. llama.cpp · IQ4_XS Decode 108 · Prefill 939 ·
  2. llama.cpp · Q4_K_M Decode 98 · Prefill 892 ·
  3. llama.cpp · Q5_K_M Decode 80 · Prefill 948 ·
  4. llama.cpp · Q8_0 Decode 66 · Prefill 724 ·
  5. vLLM · IQ4_XS Decode 82 · Prefill 946 ·
  6. vLLM · Q4_K_M Decode 73 · Prefill 855 ·
  7. vLLM · Q5_K_M Decode 75 · Prefill 843 ·
  8. vLLM · Q8_0 Decode 59 · Prefill 489 ·
  9. Ollama · IQ4_XS Decode 94 · Prefill 1095 ·
  10. Ollama · Q4_K_M Decode 87 · Prefill 997 ·
  11. Ollama · Q5_K_M Decode 89 · Prefill 724 ·
  12. Ollama · Q8_0 Decode 59 · Prefill 668 ·
  13. Strata · IQ4_XS Decode 152 · Prefill 1362 ·
  14. Strata · Q4_K_M Decode 139 · Prefill 1188 ·
  15. Strata · Q5_K_M Decode 121 · Prefill 1339 ·
  16. Strata · Q8_0 Decode 87 · Prefill 1027 ·

Fit check

Physics-based estimate from VRAM and memory bandwidth: pick a model to see which quantizations fit and the expected speed

Formula: tok/s = (bandwidth ÷ model size) × 0.55, adapted from llmfit; see run records for measured numbers

Key specs

Vendor
AMD
VRAM
48 GB
Memory bandwidth
1920 GB/s

Run records

16 configurations

ModelFrameworkQuantizationDecodeGen timeSamples
Qwen3.5-27Bllama.cppIQ4_XS108 tok/s—1
Qwen3.5-27Bllama.cppQ4_K_M98 tok/s—1
Qwen3.5-27Bllama.cppQ5_K_M80 tok/s—1
Qwen3.5-27Bllama.cppQ8_066 tok/s—1
Qwen3.5-27BvLLMIQ4_XS82 tok/s—1
Qwen3.5-27BvLLMQ4_K_M73 tok/s—1
Qwen3.5-27BvLLMQ5_K_M75 tok/s—1
Qwen3.5-27BvLLMQ8_059 tok/s—1
Qwen3.5-27BOllamaIQ4_XS94 tok/s—1
Qwen3.5-27BOllamaQ4_K_M87 tok/s—1
Qwen3.5-27BOllamaQ5_K_M89 tok/s—1
Qwen3.5-27BOllamaQ8_059 tok/s—1
Qwen3.5-27BStrataIQ4_XS152 tok/s—1
Qwen3.5-27BStrataQ4_K_M139 tok/s—1
Qwen3.5-27BStrataQ5_K_M121 tok/s—1
Qwen3.5-27BStrataQ8_087 tok/s—1