使用曦云C500 4卡推理Qwen3-32B性能极差，甚至差于使用两卡推理，但测试pcie带宽和p2p带宽没有问题，启动服务和测试的命令如下
vllm serve /home/models/Qwen3-32B --host 0.0.0.0 --port 8206 --block_size=16 --max_model_len=9120 --tensor-parallel-size 2 --gpu_memory_utilization=0.9 --no-enable-prefix-caching --async-scheduling
vllm bench serve --backend vllm --model /home/models/Qwen3-32B --host 0.0.0.0 --port 8206 --dataset-name random --random-input-len 4096 --random-output-len 1024 --ignore-eos --request-rate 40 --num-prompts 40
4卡硬件信息
=================== MetaX System Management Interface Log ===================
Timestamp : Wed May 20 10:03:09 2026
Attached GPUs : 4
+---------------------------------------------------------------------------------+
| MX-SMI 2.2.12 Kernel Mode Driver Version: 3.6.11 |
| MACA Version: 3.5.3.18 BIOS Version: 1.31.1.0 |
|------------------+-----------------+---------------------+----------------------|
| Board Name | GPU Persist-M | Bus-id | GPU-Util sGPU-M |
| Pwr:Usage/Cap | Temp Perf | Memory-Usage | GPU-State |
|==================+=================+=====================+======================|
| 0 MetaX C500 | 0 Off | 0000:81:00.0 | 0% Disabled |
| 43W / 350W | 40C P0 | 858/65536 MiB | Available |
+------------------+-----------------+---------------------+----------------------+
| 1 MetaX C500 | 1 Off | 0000:a1:00.0 | 0% Disabled |
| 40W / 350W | 37C P0 | 858/65536 MiB | Available |
+------------------+-----------------+---------------------+----------------------+
| 2 MetaX C500 | 2 Off | 0000:c1:00.0 | 0% Disabled |
| 43W / 350W | 39C P0 | 858/65536 MiB | Available |
+------------------+-----------------+---------------------+----------------------+
| 3 MetaX C500 | 3 Off | 0000:e1:00.0 | 0% Disabled |
| 43W / 350W | 38C P0 | 858/65536 MiB | Available |
+------------------+-----------------+---------------------+----------------------+
+---------------------------------------------------------------------------------+
| Process: |
| GPU PID Process Name GPU Memory |
| Usage(MiB) |
|=================================================================================|
| no process found |
使用4卡推理性能如下
============ Serving Benchmark Result ============
Successful requests: 40
Request rate configured (RPS): 40.00
Benchmark duration (s): 337.87
Total input tokens: 163734
Total generated tokens: 40960
Request throughput (req/s): 0.12
Output token throughput (tok/s): 121.23
Peak output token throughput (tok/s): 320.00
Peak concurrent requests: 40.00
Total Token throughput (tok/s): 605.83
---------------Time to First Token----------------
Mean TTFT (ms): 102917.51
Median TTFT (ms): 102972.21
P99 TTFT (ms): 196391.12
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 226.36
Median TPOT (ms): 226.71
P99 TPOT (ms): 315.54
---------------Inter-token Latency----------------
Mean ITL (ms): 226.36
Median ITL (ms): 139.24
P99 ITL (ms): 2468.52
==================================================
使用2卡推理性能如下
============ Serving Benchmark Result ============
Successful requests: 40
Request rate configured (RPS): 40.00
Benchmark duration (s): 124.06
Total input tokens: 163734
Total generated tokens: 40960
Request throughput (req/s): 0.32
Output token throughput (tok/s): 330.16
Peak output token throughput (tok/s): 800.00
Peak concurrent requests: 40.00
Total Token throughput (tok/s): 1649.94
---------------Time to First Token----------------
Mean TTFT (ms): 21047.51
Median TTFT (ms): 21005.14
P99 TTFT (ms): 40315.77
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 94.32
Median TPOT (ms): 94.41
P99 TPOT (ms): 109.96
---------------Inter-token Latency----------------
Mean ITL (ms): 94.32
Median ITL (ms): 53.47
P99 ITL (ms): 548.38
==================================================
pcie带宽如下
EFFECTIVE RAW TRANSMISSION DATA
DEVICE1 DEVICE2 TOPOLOGY SIZE(B) BANDWIDTH BANDWIDTH DELAY (us) VALIDATION
──────────────────────────────────────────────────────────────────────────────────────────────────
GPU#0 GPU#1 pcie 7516192768 101.34 GB/s 126.51 GB/s 69075.73 PASS
GPU#0 GPU#2 pcie 7516192768 101.33 GB/s 126.50 GB/s 69081.57 PASS
GPU#0 GPU#3 pcie 7516192768 101.34 GB/s 126.51 GB/s 69075.71 PASS
GPU#1 GPU#2 pcie 7516192768 101.33 GB/s 126.50 GB/s 69081.29 PASS
GPU#1 GPU#3 pcie 7516192768 101.34 GB/s 126.52 GB/s 69075.36 PASS
GPU#2 GPU#3 pcie 7516192768 101.33 GB/s 126.51 GB/s 69078.24 PASS
p2p带宽如下
EFFECTIVE RAW TRANSMISSION DATA
DEV 1 DEV 2 SIZE(B) BANDWIDTH BANDWIDTH DELAY (us) VALIDATION
──────────────────────────────────────────────────────────────────────────────────────────────────
CPU <> BOARD#0 7516192768 41.75 GB/s 52.12 GB/s 180041.45 PASS
CPU <> BOARD#1 7516192768 41.92 GB/s 52.34 GB/s 179283.75 PASS
CPU <> BOARD#2 7516192768 34.95 GB/s 43.64 GB/s 215041.24 PASS
CPU <> BOARD#3 7516192768 34.96 GB/s 43.65 GB/s 214988.19 PASS