Intel GPU 的 llama-bench 测试结果

硬件

• GPU：Intel Arc A770 + A750
• 芯片组：X570

软件

• ipex-llm： https://github.com/intel/ipex-llm/releases/download/v2.2.0-nightly/llama-cpp-ipex-llm-2.2.0b20250224-win.zip
• sycl： https://github.com/ggml-org/llama.cpp/releases/download/b4827/llama-b4827-bin-win-sycl-x64.zip
• vulkan： https://github.com/ggml-org/llama.cpp/releases/download/b4827/llama-b4827-bin-win-vulkan-x64.zip
• 模型：DeepSeek-R1-Distill-Qwen-32B-Q4_K_M.gguf

结果

llama-bench -m <模型路径>

对比用结果

4090 (572.60) + X670E, llama-b4820-bin-win-cuda-cu12.4-x64

pp512: 2291.15, tg128: 40.55

https://github.com/ProjectPhysX/OpenCL-Benchmark/releases/download/v1.8/OpenCL-Benchmark-Windows.exe

.-----------------------------------------------------------------------------. |----------------.------------------------------------------------------------| | Device ID 0 | Intel(R) Arc(TM) A770 Graphics | | Device ID 1 | Intel(R) Arc(TM) A750 Graphics | |----------------'------------------------------------------------------------| |----------------.------------------------------------------------------------| | Device ID | 0 | | Device Name | Intel(R) Arc(TM) A770 Graphics | | Device Vendor | Intel(R) Corporation | | Device Driver | 32.0.101.6559 (Windows) | | OpenCL Version | OpenCL C 3.0 | | Compute Units | 512 at 2400 MHz (4096 cores, 19.661 TFLOPs/s) | | Memory, Cache | 16255 MB VRAM, 16384 KB global / 64 KB local | | Buffer Limits | 4095 MB global, 4194296 KB constant | |----------------'------------------------------------------------------------| | Info: OpenCL C code successfully compiled. | | FP64 compute not supported | | FP32 compute 12.196 TFLOPs/s (2/3 ) | | FP16 compute 18.425 TFLOPs/s ( 1x ) | | INT64 compute 1.191 TIOPs/s (1/16) | | INT32 compute 5.687 TIOPs/s (1/4 ) | | INT16 compute 30.045 TIOPs/s ( 2x ) | | INT8 compute 29.282 TIOPs/s ( 1x ) | | Memory Bandwidth ( coalesced read ) 223.97 GB/s | | Memory Bandwidth ( coalesced write) 432.86 GB/s | | Memory Bandwidth (misaligned read ) 400.16 GB/s | | Memory Bandwidth (misaligned write) 438.62 GB/s | | PCIe Bandwidth (send ) 9.30 GB/s | | PCIe Bandwidth ( receive ) 9.00 GB/s | | PCIe Bandwidth ( bidirectional) (Gen4 x16) 9.90 GB/s | |-----------------------------------------------------------------------------| |----------------.------------------------------------------------------------| | Device ID | 1 | | Device Name | Intel(R) Arc(TM) A750 Graphics | | Device Vendor | Intel(R) Corporation | | Device Driver | 32.0.101.6559 (Windows) | | OpenCL Version | OpenCL C 3.0 | | Compute Units | 448 at 2400 MHz (3584 cores, 17.203 TFLOPs/s) | | Memory, Cache | 8095 MB VRAM, 16384 KB global / 64 KB local | | Buffer Limits | 3967 MB global, 4062248 KB constant | |----------------'------------------------------------------------------------| | Info: OpenCL C code successfully compiled. | | FP64 compute not supported | | FP32 compute 10.693 TFLOPs/s (2/3 ) | | FP16 compute 16.177 TFLOPs/s ( 1x ) | | INT64 compute 1.090 TIOPs/s (1/16) | | INT32 compute 5.043 TIOPs/s (1/3 ) | | INT16 compute 26.553 TIOPs/s ( 2x ) | | INT8 compute 26.611 TIOPs/s ( 2x ) | | Memory Bandwidth ( coalesced read ) 210.06 GB/s | | Memory Bandwidth ( coalesced write) 434.85 GB/s | | Memory Bandwidth (misaligned read ) 399.86 GB/s | | Memory Bandwidth (misaligned write) 441.22 GB/s | | PCIe Bandwidth (send ) 9.35 GB/s | | PCIe Bandwidth ( receive ) 9.04 GB/s | | PCIe Bandwidth ( bidirectional) (Gen4 x16) 9.94 GB/s | |-----------------------------------------------------------------------------| |-----------------------------------------------------------------------------| | Done. Press Enter to exit. | '-----------------------------------------------------------------------------' |----------------.------------------------------------------------------------| | Device ID | 0 | | Device Name | NVIDIA GeForce RTX 4090 | | Device Vendor | NVIDIA Corporation | | Device Driver | 572.60 (Windows) | | OpenCL Version | OpenCL C 3.0 | | Compute Units | 128 at 2535 MHz (16384 cores, 83.067 TFLOPs/s) | | Memory, Cache | 24563 MB VRAM, 3584 KB global / 48 KB local | | Buffer Limits | 6140 MB global, 64 KB constant | |----------------'------------------------------------------------------------| | Info: OpenCL C code successfully compiled. | | FP64 compute 1.401 TFLOPs/s (1/64) | | FP32 compute 85.239 TFLOPs/s ( 1x ) | | FP16 compute 88.567 TFLOPs/s ( 1x ) | | INT64 compute 4.204 TIOPs/s (1/24) | | INT32 compute 44.164 TIOPs/s (1/2 ) | | INT16 compute 38.203 TIOPs/s (1/2 ) | | INT8 compute 133.384 TIOPs/s ( 2x ) | | Memory Bandwidth ( coalesced read ) 925.72 GB/s | | Memory Bandwidth ( coalesced write) 898.38 GB/s | | Memory Bandwidth (misaligned read ) 923.73 GB/s | | Memory Bandwidth (misaligned write) 212.93 GB/s | | PCIe Bandwidth (send ) 15.66 GB/s | | PCIe Bandwidth ( receive ) 14.80 GB/s | | PCIe Bandwidth ( bidirectional) (Gen4 x16) 15.24 GB/s | |-----------------------------------------------------------------------------|