add gpu performance result

Author: sii-xinglong

blog/2025-10-29-sglang-jax.md

@@ -73,11 +73,11 @@ We benchmarked SGLang-Jax against vLLM-TPU. Full instructions are available [her
 We used `Qwen/Qwen3-32B`, TPU v6e-4, SGLang-jax (version: main-af32f095880ff676ed23eec19bc79584b5e20717), and vLLM-tpu (vllm-tpu==0.11.1).
 
 ### Results
-<img src="/images/blog/sglang_jax/performance_results.png" style="display:block; margin: auto; width: 85%;"></img>
+<img src="/images/blog/sglang_jax/tpu_performance.png" style="display:block; margin: auto; width: 85%;"></img>
 <p style="color:gray; text-align: center;">match vllm-tpu on prefill because of similar kernel optimizations. outperform vllm-tpu on decode thanks to overlap scheduler. </p>
 
-Todo:
-- (optional) show some TPUs vs. GPUs.
+<img src="/images/blog/sglang_jax/gpu_performance.png" style="display:block; margin: auto; width: 85%;"></img>
+<p style="color:gray; text-align: center;">the TPU setup achieves lower latency (TTFT and ITL) and higher input throughput across various batch sizes</p>
 
 ## Usage
 
@@ -156,7 +156,7 @@ The community is working with Google Cloud team and multiple partners on the fol
    - Multi-LoRA batching
 
 ## Acknowledgments
-**SGLang-jax team**: sii-xinglong, jimoosciuc, Prayer, aolemila, JamesBrianD, zkkython, neo, leos, pathfinder-pf, Ying Sheng, Hongzhen Chen, Jiacheng Yang, Ke Bao
+**SGLang-jax team**: sii-xinglong, jimoosciuc, Prayer, aolemila, JamesBrianD, zkkython, neo, leos, pathfinder-pf, Ying Sheng, Hongzhen Chen, Jiacheng Yang, Ke Bao, Qinghan Chen
 
 **Google**: Google Cloud Team