基本配置

Specifications
Xe-cores20
Graphics Clock2670 MHz
Memory12 GB GDDR6
Graphics Memory Interface192 bit
Graphics Memory Bandwidth456 GB/s
GPU Peak TOPS (Int8)233
TBP190 W

系统环境

kernal: 6.12.7
mesa: 24.3.2
os: fedora 41
egpu: m2 to oculink

PCIe Gen 1x1

lspci 对于 Intel Arc 显卡总是显示 PCIe 1x1 的速度
这似乎是预期行为，并不是实际的速度，Troubleshooting， Intel Support Forums

具体的连接速度可以用 xpumanager 来测试

Resizable BAR

Intel 说明 Arc 显卡需要开启 Resizable BAR 才能完全发挥性能
查了下资料，Base Address Register 用于把 pcie 设备的资源映射到系统内存地址，内部有一套同步逻辑

台式主板一般有直接的选项开启
笔记本和mini主机需要把分配给核显的内存开到 4G 以上

$ sudo lspci -s 03:00.0 -vv | grep -i bar 
     Capabilities: [420 v1] Physical Resizable BAR 
         BAR 2: current size: 16GB, supported: 256MB 512MB 1GB 2GB 4GB 8GB 16GB

AI 应用

llama.cpp

支持 SYCL 和 Vulkan 后端，我写好了 Dockerfile.
SYCL 版本需要安装 intel oneapi，然后在构建时挂载进容器
llama-bench -ngl 99 结果(统计参考)：

  • SYCL

    modelsizeparamsbackendngltestt/s
    llama 7B Q4_03.56 GiB6.74 BSYCL99pp5121982.69 ± 4.02
    llama 7B Q4_03.56 GiB6.74 BSYCL99tg12834.83 ± 0.11

  • Vulkan

    ggml_vulkan: 0 = Intel(R) Graphics (BMG G21) (Intel open-source Mesa driver) | uma: 0 | fp16: 1 | warp size: 32 | matrix cores: none
    modelsizeparamsbackendngltestt/s
    llama 7B Q4_03.56 GiB6.74 BVulkan99pp512175.56 ± 2.65
    llama 7B Q4_03.56 GiB6.74 BVulkan99tg12844.12 ± 0.09

TODO

游戏

TODO

性能测试

GravityMark

统计网站

  • Vulkan 1920x1080 Asteroids: 200000 Score: 24896
  • Opengl 1920x1080 Asteroids: 200000 Score: 20477

clpeak

统计网站

Single-precision compute 和 4060 差不多
Half-precision compute 网站没有数据，看下面 vkpeak, 和 4070 super 差不多
Integer compute 差不多 4060 的一半1

$ ./clpeak 

Platform: Intel(R) OpenCL Graphics
  Device: Intel(R) Graphics [0xe20b]
    Driver version  : 24.35.30872.32 (Linux x64)
    Compute units   : 160
    Clock frequency : 2850 MHz

    Global memory bandwidth (GBPS)
      float   : 417.71
      float2  : 429.27
      float4  : 434.78
      float8  : 446.30
      float16 : 449.03

    Single-precision compute (GFLOPS)
      float   : 14083.21
      float2  : 14401.35
      float4  : 14096.05
      float8  : 12096.30
      float16 : 13596.31

    Half-precision compute (GFLOPS)
      half   : 24831.84
      half2  : 28076.27
      half4  : 28199.86
      half8  : 27916.44
      half16 : 27768.14

    Double-precision compute (GFLOPS)
      double   : 891.12
      double2  : 893.37
      double4  : 898.90
      double8  : 888.26
      double16 : 842.54

    Integer compute (GIOPS)
      int   : 4433.77
      int2  : 4461.17
      int4  : 4456.32
      int8  : 4364.93
      int16 : 4192.12

    Integer compute Fast 24bit (GIOPS)
      int   : 4473.72
      int2  : 4448.66
      int4  : 4470.87
      int8  : 4379.65
      int16 : 4225.50

    Integer char (8bit) compute (GIOPS)
      char   : 21487.19
      char2  : 26128.29
      char4  : 25867.58
      char8  : 25227.17
      char16 : 24600.19

    Integer short (16bit) compute (GIOPS)
      short   : 20924.35
      short2  : 26189.49
      short4  : 25835.04
      short8  : 25127.30
      short16 : 23874.53

    Transfer bandwidth (GBPS)
      enqueueWriteBuffer              : 6.77
      enqueueReadBuffer               : 6.71
      enqueueWriteBuffer non-blocking : 7.08
      enqueueReadBuffer non-blocking  : 7.01
      enqueueMapBuffer(for read)      : 6.97
        memcpy from mapped ptr        : 26.67
      enqueueUnmap(after write)       : 7.20
        memcpy to mapped ptr          : 26.41

    Kernel launch latency : 49.15 us

vkpeak

统计网站

fp32 和 4060 差不多
int32 特别差
其他的居然能跑到和 4070 super 差不多

$ ./vkpeak 0 
device       = Intel(R) Graphics (BMG G21)

fp32-scalar  = 7277.51 GFLOPS
fp32-vec4    = 10706.95 GFLOPS

fp16-scalar  = 21068.69 GFLOPS
fp16-vec4    = 23754.69 GFLOPS
fp16-matrix  = 0.00 GFLOPS

fp64-scalar  = 799.67 GFLOPS
fp64-vec4    = 779.47 GFLOPS

int32-scalar = 3024.92 GIOPS
int32-vec4   = 3102.06 GIOPS

int16-scalar = 10801.67 GIOPS
int16-vec4   = 13525.01 GIOPS

VkFFT

统计网站

  • single precision
$ ./VkFFT_TestSuite -vkfft 0
0 - VkFFT FFT + iFFT C2C benchmark 1D batched in single precision
VkFFT System: 3 8x16777216 Buffer: 1024 MB avg_time_per_step: 10.984 ms std_error: 0.054 num_iter: 3 benchmark: 95462 bandwidth: 364.2
VkFFT System: 4 16x8388608 Buffer: 1024 MB avg_time_per_step: 11.302 ms std_error: 0.378 num_iter: 3 benchmark: 92774 bandwidth: 353.9
VkFFT System: 5 32x4194304 Buffer: 1024 MB avg_time_per_step: 11.164 ms std_error: 0.332 num_iter: 3 benchmark: 93921 bandwidth: 358.3
VkFFT System: 6 64x2097152 Buffer: 1024 MB avg_time_per_step: 10.819 ms std_error: 0.024 num_iter: 3 benchmark: 96918 bandwidth: 369.7
VkFFT System: 7 128x1048576 Buffer: 1024 MB avg_time_per_step: 10.816 ms std_error: 0.022 num_iter: 3 benchmark: 96946 bandwidth: 369.8
VkFFT System: 8 256x524288 Buffer: 1024 MB avg_time_per_step: 10.918 ms std_error: 0.016 num_iter: 3 benchmark: 96037 bandwidth: 366.4
VkFFT System: 9 512x262144 Buffer: 1024 MB avg_time_per_step: 10.963 ms std_error: 0.045 num_iter: 3 benchmark: 95645 bandwidth: 364.9
VkFFT System: 10 1024x131072 Buffer: 1024 MB avg_time_per_step: 10.990 ms std_error: 0.052 num_iter: 3 benchmark: 95414 bandwidth: 364.0
VkFFT System: 11 2048x65536 Buffer: 1024 MB avg_time_per_step: 11.018 ms std_error: 0.067 num_iter: 3 benchmark: 95172 bandwidth: 363.1
VkFFT System: 12 4096x32768 Buffer: 1024 MB avg_time_per_step: 10.981 ms std_error: 0.080 num_iter: 3 benchmark: 95486 bandwidth: 364.3
VkFFT System: 13 8192x16384 Buffer: 1024 MB avg_time_per_step: 11.638 ms std_error: 1.004 num_iter: 3 benchmark: 90095 bandwidth: 343.7
  • half precision
$ ./VkFFT_TestSuite -vkfft 2
2 - VkFFT FFT + iFFT C2C benchmark 1D batched in half precision
VkFFT System: 3 8x16777216 Buffer: 512 MB avg_time_per_step: 5.755 ms std_error: 0.054 num_iter: 2 benchmark: 182207 bandwidth: 347.5
VkFFT System: 4 16x8388608 Buffer: 512 MB avg_time_per_step: 5.855 ms std_error: 0.018 num_iter: 2 benchmark: 179090 bandwidth: 341.6
VkFFT System: 5 32x4194304 Buffer: 512 MB avg_time_per_step: 6.394 ms std_error: 0.066 num_iter: 2 benchmark: 163989 bandwidth: 312.8
VkFFT System: 6 64x2097152 Buffer: 512 MB avg_time_per_step: 6.731 ms std_error: 0.127 num_iter: 2 benchmark: 155790 bandwidth: 297.1
VkFFT System: 7 128x1048576 Buffer: 512 MB avg_time_per_step: 7.127 ms std_error: 0.059 num_iter: 2 benchmark: 147137 bandwidth: 280.6
VkFFT System: 8 256x524288 Buffer: 512 MB avg_time_per_step: 5.517 ms std_error: 0.100 num_iter: 2 benchmark: 190062 bandwidth: 362.5
VkFFT System: 9 512x262144 Buffer: 512 MB avg_time_per_step: 5.403 ms std_error: 0.004 num_iter: 2 benchmark: 194084 bandwidth: 370.2
VkFFT System: 10 1024x131072 Buffer: 512 MB avg_time_per_step: 5.575 ms std_error: 0.168 num_iter: 2 benchmark: 188085 bandwidth: 358.7
VkFFT System: 11 2048x65536 Buffer: 512 MB avg_time_per_step: 5.560 ms std_error: 0.126 num_iter: 2 benchmark: 188598 bandwidth: 359.7
VkFFT System: 12 4096x32768 Buffer: 512 MB avg_time_per_step: 5.590 ms std_error: 0.108 num_iter: 2 benchmark: 187580 bandwidth: 357.8
VkFFT System: 13 8192x16384 Buffer: 512 MB avg_time_per_step: 10.461 ms std_error: 1.182 num_iter: 2 benchmark: 100241 bandwidth: 191.2