Update (#190)

* update

* update

* update amd day0

Author: hnyls2002

blog/2025-08-27-gpt-oss.md

@@ -26,21 +26,27 @@ To show the impact of our optimizations, we benchmarked SGLang across a range of
 
 ##### Low-Latency Performance (Batch Size = 1)
 
-For latency-sensitive applications, we measured single-batch decode throughput across NVIDIA and AMD GPUs, showcasing excellent performance.
+For latency-sensitive applications, we measured single-batch decode throughput across B200 and H100 GPUs, showcasing excellent performance.
 
-| Hardware / Precision | NVIDIA B200  | NVIDIA H100  | AMD MI350    |
-| -------------------- | ------------ | ------------ | ------------ |
-| MXFP4                | 416.02 tok/s | 318.53 tok/s | 200.84 tok/s |
-| BF16                 | 315.63 tok/s | 293.12 tok/s | 220.06 tok/s |
+| Hardware / Precision | NVIDIA B200  | NVIDIA H100  |
+| -------------------- | ------------ | ------------ |
+| MXFP4                | 416.02 tok/s | 318.53 tok/s |
+| BF16                 | 315.63 tok/s | 293.12 tok/s |
 
 <span style="color: grey; font-size: 12px;">
-B200 was tested with TP=4, H100 with TP=8 and triton attention, and MI350 with TP=8 and triton backend.
+B200 was tested with TP=4, H100 was tested with TP=8 and triton attention.
 </span>
 
 ##### High-Throughput Performance (Batch Size = 32)
 
 For high-throughput applications, SGLang delivers significant performance gains over our initial Day 0 support and have shown great performance on both prefill and decode on different hardwares.
 
+<!-- grey text -->
+
+<span style="color: grey; font-size: 12px;">
+The results of AMD MI350 were tested with triton backend which is not fully optimized yet, and more optimizations with AMD AITER will be released soon.
+</span>
+
 <img src="/images/blog/gpt_oss/combined_prefill_performance.svg" alt="combined_prefill_performance.svg" style="display:block; margin-left: auto; margin-right: auto; width: 75%"></img>
 
 <img src="/images/blog/gpt_oss/combined_decode_performance.svg" alt="combined_decode_performance.svg" style="display:block; margin-left: auto; margin-right: auto; width: 75%"></img>
@@ -127,6 +133,6 @@ print(response.output_text)
 
 None of the Day-0 support or the subsequent optimizations would have been possible without the collective effort of the SGLang community. Shout-out to the SGLang team, SpecForge team, FlashInfer team, Oracle team, Eigen AI team, NVIDIA team and AMD team for pushing this forward together!
 
-We will continue pushing the boundaries of LLM inference. On our roadmap are further explorations into SWA (Sliding Window Attention) optimizations, along with new advances in speculative decoding, to deliver even greater performance gains.
+We will continue pushing the boundaries of LLM inference. On our roadmap are further explorations into SWA (Sliding Window Attention) optimizations, AMD AITER integration, along with new advances in speculative decoding, to deliver even greater performance gains.
 
 We invite you to try the latest version of SGLang and share your feedback. Thank you for being an essential part of this journey!