Intel Arc B580

GPU

77

Run records

14

Models covered

4

Frameworks covered

L0 Self-reported × 1 · L1 Reproduced × 1 · L2 Cross-framework verified × 75

Evidence mix

Typical measured performance

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

  1. llama.cpp · IQ3_XS Decode 70.76 · Prefill 300.15 ·
  2. Ollama · IQ3_S Decode 65.12 · Prefill 253.41 ·
  3. Ollama · Q2_K Decode 60.73 · Prefill 314.93 ·
  4. llama.cpp · IQ2_M Decode 164.15 · Prefill 679.38 ·
  5. Ollama · Q3_K_L Decode 127.75 · Prefill 669.2 ·
  6. llama.cpp · IQ2_XS Decode 84.02 · Prefill 284.07 ·
  7. llama.cpp · IQ4_XS Decode 64.83 · Prefill 327.91 ·
  8. Ollama · IQ2_M Decode 78.85 · Prefill 313.28 ·
  9. Ollama · Q8_0 Decode 37.78 · Prefill 141.17 ·
  10. llama.cpp · IQ3_S Decode 138.57 · Prefill 710.44 ·
  11. llama.cpp · IQ3_XS Decode 131.95 · Prefill 488.13 ·
  12. Ollama · Q4_K_S Decode 101.11 · Prefill 350.9 ·
  13. Ollama · Q8_0 Decode 66.71 · Prefill 216.71 ·
  14. llama.cpp · IQ2_M Decode 59.78 · Prefill 260.58 ·
  15. Ollama · IQ1_S Decode 60.08 · Prefill 208.14 ·
  16. llama.cpp · IQ3_S Decode 58.55 · Prefill 274.06 ·
  17. llama.cpp · Q3_K_S Decode 65.46 · Prefill 319.04 ·
  18. Ollama · IQ2_XS Decode 70.97 · Prefill 264.22 ·
  19. llama.cpp · Q3_K_M Decode 57.84 · Prefill 315.13 ·
  20. Ollama · Q2_K Decode 67.02 · Prefill 263.91 ·
  21. llama.cpp · Q2_K Decode 169.62 · Prefill 813.51 ·
  22. llama.cpp · Q5_K_S Decode 120.67 · Prefill 490.78 ·
  23. Ollama · FP16 Decode 62.55 · Prefill 296.07 ·
  24. llama.cpp · Q3_K_S Decode 74.37 · Prefill 373.85 ·
  25. Strata · Q6_K Decode 60.78 · Prefill 199.46 ·
  26. Ollama · Q3_K_M Decode 110.96 · Prefill 517.6 ·
  27. Strata · Q5_K_M Decode 136.37 · Prefill 553.15 ·
  28. vLLM · IQ1_S Decode 66.31 · Prefill 217.11 ·
  29. vLLM · Q3_K_M Decode 56.47 · Prefill 307.9 ·
  30. Strata · IQ2_M Decode 88.38 · Prefill 340.69 ·
  31. Strata · Q3_K_M Decode 71.01 · Prefill 283.66 ·
  32. Ollama · IQ2_M Decode 67.87 · Prefill 259.33 ·
  33. llama.cpp · IQ3_S Decode 60.72 · Prefill 260.85 ·
  34. vLLM · IQ1_S Decode 85.09 · Prefill 376.72 ·
  35. vLLM · Q2_K Decode 76.78 · Prefill 352.63 ·
  36. vLLM · IQ3_XS Decode 157.7 · Prefill 718.89 ·
  37. vLLM · Q4_K_M Decode 135.38 · Prefill 477.15 ·
  38. Strata · FP16 Decode 92 · Prefill 368.95 ·
  39. Strata · Q4_K_M Decode 163.66 · Prefill 710.01 ·
  40. llama.cpp · IQ3_S Decode 47.04 · Prefill 172.07 ·
  41. llama.cpp · Q3_K_L Decode 43.5 · Prefill 143.66 ·
  42. Strata · IQ2_XS Decode 78.95 · Prefill 342.16 ·
  43. llama.cpp · IQ1_S Decode 98.24 · Prefill 506.66 ·
  44. llama.cpp · Q3_K_L Decode 70.52 · Prefill 317.93 ·
  45. vLLM · Q2_K Decode 89.29 · Prefill 348.31 ·
  46. Strata · IQ3_S Decode 101.32 · Prefill 490.59 ·
  47. llama.cpp · Q4_K_M Decode 117.19 · Prefill 568.87 ·
  48. vLLM · Q4_K_S Decode 137.77 · Prefill 674.05 ·
  49. llama.cpp · IQ2_XS Decode 61.39 · Prefill 302.91 ·
  50. llama.cpp · IQ3_XS Decode 56.63 · Prefill 277.66 ·

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
Intel
VRAM
12 GB
Memory bandwidth
456 GB/s

Run records

50 configurations

ModelFrameworkQuantizationDecodeGen timeSamples
gemma-4-E4B-itllama.cppIQ3_XS70.8 tok/s—1
gemma-4-E4B-itOllamaIQ3_S65.1 tok/s—1
Ministral-3-8B-Instruct-2512OllamaQ2_K60.7 tok/s—1
iFlytek Spark-X2.5-4Bllama.cppIQ2_M164.2 tok/s—1
iFlytek Spark-X2.5-4BOllamaQ3_K_L127.8 tok/s—1
Qwen3.5-4Bllama.cppIQ2_XS84 tok/s—1
Qwen3.5-4Bllama.cppIQ4_XS64.8 tok/s—1
Qwen3.5-4BOllamaIQ2_M78.8 tok/s—1
Qwen3.5-4BOllamaQ8_037.8 tok/s—1
Qwen3.5-0.8Bllama.cppIQ3_S138.6 tok/s—1
Qwen3.5-0.8Bllama.cppIQ3_XS131.9 tok/s—1
Qwen3.5-0.8BOllamaQ4_K_S101.1 tok/s—1
Qwen3.5-0.8BOllamaQ8_066.7 tok/s—1
Qwen3.8-9Bllama.cppIQ2_M59.8 tok/s—2
Qwen3.8-9BOllamaIQ1_S60.1 tok/s—1
Llama-3.1-8B-Instructllama.cppIQ3_S58.5 tok/s—1
Llama-3.1-8B-Instructllama.cppQ3_K_S65.5 tok/s—1
Llama-3.1-8B-InstructOllamaIQ2_XS71 tok/s—1
Llama-3-8B-Instructllama.cppQ3_K_M57.8 tok/s—1
Llama-3-8B-InstructOllamaQ2_K67 tok/s—1
Llama-3.2-3B-Instructllama.cppQ2_K169.6 tok/s—1
Llama-3.2-3B-Instructllama.cppQ5_K_S120.7 tok/s—1
Llama-3.2-3B-InstructOllamaFP1662.5 tok/s—1
Qwen3.5-4Bllama.cppQ3_K_S74.4 tok/s—1
Qwen3.5-4BStrataQ6_K60.8 tok/s—1
Qwen3.5-0.8BOllamaQ3_K_M111 tok/s—1
Qwen3.5-0.8BStrataQ5_K_M136.4 tok/s—1
Qwen3.8-9BvLLMIQ1_S66.3 tok/s—1
Qwen3-8BvLLMQ3_K_M56.5 tok/s—1
Qwen3-8BStrataIQ2_M88.4 tok/s—1
Qwen3-8BStrataQ3_K_M71 tok/s—1
Llama-3.1-8B-InstructOllamaIQ2_M67.9 tok/s—1
Llama-3-8B-Instructllama.cppIQ3_S60.7 tok/s—1
Llama-3-8B-InstructvLLMIQ1_S85.1 tok/s—1
Llama-3-8B-InstructvLLMQ2_K76.8 tok/s—1
Llama-3.2-3B-InstructvLLMIQ3_XS157.7 tok/s—1
Llama-3.2-3B-InstructvLLMQ4_K_M135.4 tok/s—1
Llama-3.2-3B-InstructStrataFP1692 tok/s—1
Llama-3.2-3B-InstructStrataQ4_K_M163.7 tok/s—1
Kimi-K3llama.cppIQ3_S47 tok/s—1
Kimi-K3llama.cppQ3_K_L43.5 tok/s—1
Kimi-K3StrataIQ2_XS79 tok/s—1
Qwen3.5-4Bllama.cppIQ1_S98.2 tok/s—1
Qwen3.5-4Bllama.cppQ3_K_L70.5 tok/s—1
Qwen3.5-4BvLLMQ2_K89.3 tok/s—1
Qwen3.5-4BStrataIQ3_S101.3 tok/s—1
Qwen3.5-0.8Bllama.cppQ4_K_M117.2 tok/s—1
Qwen3.5-0.8BvLLMQ4_K_S137.8 tok/s—1
Qwen3.8-9Bllama.cppIQ2_XS61.4 tok/s—1
Qwen3.8-9Bllama.cppIQ3_XS56.6 tok/s—1