Back to Benchmarks

Qwen/Qwen3-235B-A22B

A full-stack benchmark for high-throughput MoE inference on Multi-GPUs.

+138.7% token/s · 58.7% Cost Savings over the baseline

MODEL OVERVIEW

Qwen3-235B-A22B Overview

Qwen3-235B-A22B is a sparse Mixture-of-Experts causal language model with 235B total parameters and 22B active per token. It uses 128 experts, with 8 active per token, a 94-layer architecture, a 4,096 hidden size, and grouped-query attention with 64 query heads and 4 KV heads. The model supports a native context length of 32,768 tokens, which can be extended up to 131,072 tokens using YaRN.

Optimized Inference Environment

A single-node vLLM deployment tuned for high-throughput serving of Qwen3-235B-A22B.

HARDWARE

8 × NVIDIA H100-80GB · single node

VLLM

v0.20.0

PARALLELISM

Tensor parallel 8 · mp distributed executor

FEATURES

Chunked prefill · prefix caching · Triton MoE backend

Workload Profile

A controlled synthetic workload designed to reflect high-concurrency serving conditions.

FRAMEWORK

vllm bench serve

DATASET

random (synthetic) · range ratio 0.8

INPUT LENGTH

409–3,686 tokens · nominal 2,048 · uniform

OUTPUT LENGTH

51–460 tokens · nominal 256 · uniform

PROMPTS MEASURED

1,000 · warm-ups excluded: 10

REQUEST RATE

inf · max concurrency is the only throttle

TESTED CONCURRENCY

512

Baseline and Optimized Serving Configurations

BASELINE

vllm serve Qwen/Qwen3-235B-A22B \

--host 0.0.0.0 --port 8000 \

--trust-remote-code \

--tensor-parallel-size 8

OPTIMIZED

export VLLM_WORKER_MULTIPROC_METHOD=spawn

export CUDA_DEVICE_MAX_CONNECTIONS=1


vllm serve Qwen/Qwen3-235B-A22B \

--tensor-parallel-size 8 \

--gpu-memory-utilization 0.90 \

--max-num-batched-tokens 8192 \

--max-num-seqs 512 \

--enable-chunked-prefill \

--enable-prefix-caching \

--max-model-len 8192 \

--distributed-executor-backend mp \

--moe-backend triton

Benchmark Results

Baseline and optimized serving results for the multi-GPU benchmark configuration.

Metric

Baseline

Optimized

Output tok/s

759.0

1,812.0

Request throughput

2.92 req/s

6.98 req/s

TTFT mean

4,304.4 ms

3,786.8 ms

TTFT p99

16,379.8 ms

10,526.9 ms

TPOT mean

146.3 ms

53.9 ms

TPOT p99

195.1 ms

72.6 ms

ITL mean

144.3 ms

53.2 ms

ITL p99

535.7 ms

240.6 ms

E2E mean

41,766.3 ms

17,589.2 ms

E2E median

40,450.2 ms

17,168.5 ms

Duration

342.1 s

143.3 s

Optimization/Tuning Changes

Tuning changes applied to increase usable cache capacity, batch depth, and MoE execution efficiency.

max-model-len

Capped at 8,192 to convert unused context into KV cache slots.

gpu-memory-utilization = 0.90

Reserves 90% of GPU memory for serving capacity.

max-num-seqs = 512

Allows deeper request batching.

Chunked prefill + prefix caching

Enabled to improve reuse and reduce prefill pressure.

Triton MoE backend + mp distributed executor

Selected for efficient expert execution across the eight-way tensor-parallel deployment.

Environment

VLLM_WORKER_MULTIPROC_METHOD=spawn · CUDA_DEVICE_MAX_CONNECTIONS=1

Cost Savings

At the same GPU-hour rate, higher optimized throughput reduces the effective cost of each generated token.

THROUGHPUT GAIN

+138.7%

1,812 vs 759 output tok/s

ESTIMATED SAVINGS

58.1%

same GPU-hour rate

Calculation: 1 − (759 ÷ 1,812) = 58.1% estimated savings. Duration corroborates the estimate: 342.1s → 143.3s.