[Performance]: Unexpected: B200 GPU Performance Similar to H200 for Qwen/QwQ-32B, Expected B200 to be Significantly Faster

[awterman]

Proposal to improve performance

We are observing that the B200 GPU is performing similarly to the H200 GPU when running inference with the Qwen/QwQ-32B model using vLLM. We expect the B200 to have significantly better performance.

Hardware Information:

• CPU: 192 x vCPU
• Memory: 1585 GB

Benchmark Script:

vllm serve Qwen/QwQ-32B

python3 ./vllm/benchmarks/benchmark_serving.py \
        --backend vllm \
        --model Qwen/QwQ-32B \
        --max-concurrency 1 \
        --base-url http://127.0.0.1:8000 \
        --endpoint /v1/completions \
        --dataset-name sharegpt \
        --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json \
        --num-prompts 10

Performance Metrics:

H200 (8x H200 QwQ-32B elaich/simple-vllm-launcher):

============ Serving Benchmark Result ============
Successful requests:                     10
Benchmark duration (s):                  23.10
Total input tokens:                      1374
Total generated tokens:                  2663
Request throughput (req/s):              0.43
Output token throughput (tok/s):         115.31
Total Token throughput (tok/s):          174.80
---------------Time to First Token----------------
Mean TTFT (ms):                          585.38
Median TTFT (ms):                        594.23
P99 TTFT (ms):                           615.97
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          6.70
Median TPOT (ms):                        6.49
P99 TPOT (ms):                           8.23
---------------Inter-token Latency----------------
Mean ITL (ms):                           6.49
Median ITL (ms):                         6.43
P99 ITL (ms):                            13.49
==================================================

B200 (8x B200 QwQ-32B datura-vllm-base simple-vllm-launcher):

============ Serving Benchmark Result ============
Successful requests:                     10
Benchmark duration (s):                  23.33
Total input tokens:                      1374
Total generated tokens:                  2663
Request throughput (req/s):              0.43
Output token throughput (tok/s):         114.16
Total Token throughput (tok/s):          173.07
---------------Time to First Token----------------
Mean TTFT (ms):                          342.74
Median TTFT (ms):                        329.84
P99 TTFT (ms):                           487.91
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          7.46
Median TPOT (ms):                        7.51
P99 TPOT (ms):                           7.59
---------------Inter-token Latency----------------
Mean ITL (ms):                           7.50
Median ITL (ms):                         7.51
P99 ITL (ms):                            13.53
==================================================

Expected Behavior:

The B200 GPU should exhibit significantly higher throughput (tok/s) and lower latencies (TTFT, TPOT, ITL) compared to the H200 GPU.

Actual Behavior:

The performance metrics for both GPUs are very similar, with the B200 showing only marginal improvements in TTFT, while TPOT and ITL are slightly worse.

Steps to Reproduce:

1. Set up vLLM on a machine with 8x H200 GPUs.
2. Run the benchmark script provided above.
3. Record the performance metrics.
4. Set up vLLM on a machine with 8x B200 GPUs.
5. Run the benchmark script provided above.
6. Compare the performance metrics.

Please let us know if any further information is required.

Report of performance regression

No response

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

The output of `python collect_env.py`

Collecting environment information...

    System Info

==============================
OS : Ubuntu 22.04.5 LTS (x86_64)
GCC version : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version : Could not collect
CMake version : Could not collect
Libc version : glibc-2.35

==============================
PyTorch Info

PyTorch version : 2.7.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A

==============================
Python Environment

Python version : 3.11.12 | packaged by conda-forge | (main, Apr 10 2025, 22:23:25) [GCC 13.3.0] (64-bit runtime)
Python platform : Linux-5.15.0-138-generic-x86_64-with-glibc2.35

==============================
CUDA / GPU Info

Is CUDA available : True
CUDA runtime version : Could not collect
CUDA_MODULE_LOADING set to : LAZY
GPU models and configuration :
GPU 0: NVIDIA B200
GPU 1: NVIDIA B200
GPU 2: NVIDIA B200
GPU 3: NVIDIA B200
GPU 4: NVIDIA B200
GPU 5: NVIDIA B200
GPU 6: NVIDIA B200
GPU 7: NVIDIA B200

Nvidia driver version : 570.133.20
cuDNN version : Could not collect
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True

==============================
CPU Info

Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 192
On-line CPU(s) list: 0-191
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) 6972P
CPU family: 6
Model: 173
Thread(s) per core: 1
Core(s) per socket: 96
Socket(s): 2
Stepping: 1
BogoMIPS: 4800.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 9 MiB (192 instances)
L1i cache: 12 MiB (192 instances)
L2 cache: 384 MiB (192 instances)
L3 cache: 960 MiB (2 instances)
NUMA node(s): 6
NUMA node0 CPU(s): 0-31
NUMA node1 CPU(s): 32-63
NUMA node2 CPU(s): 64-95
NUMA node3 CPU(s): 96-127
NUMA node4 CPU(s): 128-159
NUMA node5 CPU(s): 160-191
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

==============================
Versions of relevant libraries

[pip3] numpy==2.2.5
[pip3] nvidia-cublas-cu12==12.8.3.14
[pip3] nvidia-cuda-cupti-cu12==12.8.57
[pip3] nvidia-cuda-nvrtc-cu12==12.8.61
[pip3] nvidia-cuda-runtime-cu12==12.8.57
[pip3] nvidia-cudnn-cu12==9.7.1.26
[pip3] nvidia-cufft-cu12==11.3.3.41
[pip3] nvidia-cufile-cu12==1.13.0.11
[pip3] nvidia-curand-cu12==10.3.9.55
[pip3] nvidia-cusolver-cu12==11.7.2.55
[pip3] nvidia-cusparse-cu12==12.5.7.53
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-ml-py==12.575.51
[pip3] nvidia-nccl-cu12==2.26.2
[pip3] nvidia-nvjitlink-cu12==12.8.61
[pip3] nvidia-nvtx-cu12==12.8.55
[pip3] optree==0.15.0
[pip3] pynvml==12.0.0
[pip3] pyzmq==26.4.0
[pip3] torch==2.7.0+cu128
[pip3] torchaudio==2.7.0+cu128
[pip3] torchelastic==0.2.2
[pip3] torchvision==0.22.0+cu128
[pip3] transformers==4.51.3
[pip3] triton==3.3.0
[conda] Could not collect

==============================
vLLM Info

ROCM Version : Could not collect
Neuron SDK Version : N/A
vLLM Version : 0.8.5.dev649+g0189a65a2 (git sha: 0189a65)
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 NIC7 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 PIX NODE SYS SYS SYS SYS SYS SYS 0-31 0 N/A
GPU1 NV18 X NV18 NV18 NV18 NV18 NV18 NV18 NODE PIX SYS SYS SYS SYS SYS SYS 0-31 0 N/A
GPU2 NV18 NV18 X NV18 NV18 NV18 NV18 NV18 SYS SYS PIX NODE SYS SYS SYS SYS 64-95 2 N/A
GPU3 NV18 NV18 NV18 X NV18 NV18 NV18 NV18 SYS SYS NODE PIX SYS SYS SYS SYS 64-95 2 N/A
GPU4 NV18 NV18 NV18 NV18 X NV18 NV18 NV18 SYS SYS SYS SYS PIX NODE SYS SYS 96-127 3 N/A
GPU5 NV18 NV18 NV18 NV18 NV18 X NV18 NV18 SYS SYS SYS SYS NODE PIX SYS SYS 96-127 3 N/A
GPU6 NV18 NV18 NV18 NV18 NV18 NV18 X NV18 SYS SYS SYS SYS SYS SYS PIX NODE 160-191 5 N/A
GPU7 NV18 NV18 NV18 NV18 NV18 NV18 NV18 X SYS SYS SYS SYS SYS SYS NODE PIX 160-191 5 N/A
NIC0 PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS SYS SYS SYS
NIC1 NODE PIX SYS SYS SYS SYS SYS SYS NODE X SYS SYS SYS SYS SYS SYS
NIC2 SYS SYS PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS SYS
NIC3 SYS SYS NODE PIX SYS SYS SYS SYS SYS SYS NODE X SYS SYS SYS SYS
NIC4 SYS SYS SYS SYS PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS
NIC5 SYS SYS SYS SYS NODE PIX SYS SYS SYS SYS SYS SYS NODE X SYS SYS
NIC6 SYS SYS SYS SYS SYS SYS PIX NODE SYS SYS SYS SYS SYS SYS X NODE
NIC7 SYS SYS SYS SYS SYS SYS NODE PIX SYS SYS SYS SYS SYS SYS NODE X

Legend:

X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks

NIC Legend:

NIC0: mlx5_0
NIC1: mlx5_1
NIC2: mlx5_2
NIC3: mlx5_3
NIC4: mlx5_4
NIC5: mlx5_9
NIC6: mlx5_10
NIC7: mlx5_11

==============================
Environment Variables

NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.