Delivering 1.5M TPS Inference on NVIDIA GB200 NVL72, NVIDIA Accelerates OpenAI gpt-oss Models From Cloud to Edge

NVIDIA and OpenAI began pushing the boundaries of AI with the launch of NVIDIA DGX back in 2016. The collaborative AI innovation continues with the OpenAI gpt-oss-20b and gpt-oss-120b launch. NVIDIA has optimized both new open-weight models for accelerated inference performance on NVIDIA Blackwell architecture, delivering up to 1.5 million tokens per second (TPS) on an NVIDIA GB200 NVL72 system.

The gpt-oss models are text-reasoning LLMs with chain-of-thought and tool-calling capabilities using the popular mixture of experts (MoE) architecture with SwigGLU activations. The attention layers use RoPE with 128k context, alternating between full context and a sliding 128-token window. The models are released in FP4 precision, which fits on a single 80 GB data center GPU and is natively supported by Blackwell.

The models were trained on NVIDIA H100 Tensor Core GPUs, with gpt-oss-120b requiring over 2.1 million hours and gpt-oss-20b about 10x less. NVIDIA worked with several top open-source frameworks such as Hugging Face Transformers, Ollama, and vLLM, in addition to NVIDIA TensorRT-LLM for optimized kernels and model enhancements. This blog post showcases how NVIDIA has integrated gpt-oss across the software platform to meet developers' needs.

Model name Transformer Blocks Total Parameters Active Params per Token # of Experts Active Experts per Token Input Context Length

gpt-oss-20b 24 20B 3.6B 32 4 128K

gpt-oss-120b 36 117B 5.1B 128 4 128K

Table 1. OpenAI gpt-oss-20b and gpt-oss-120b model specifications, including total parameters, active parameters, number of experts, and input context length NVIDIA also worked with OpenAI and the community to maximize performance, adding features such as:

TensorRT-LLM Gen for attention prefill, attention decode, and MoE low-latency on Blackwell.

CUTLASS MoE kernels on Blackwell.

XQA kernel for specialized attention on Hopper.

Optimized attention and MoE routing kernels are available through the FlashInfer kernel-serving library for LLMs.

OpenAI Triton kernel MoE support, which is used in both TensorRT-LLM and vLLM.

Deploy using vLLM In collaboration with vLLM, NVIDIA worked together to verify accuracy while also analyzing and optimizing performance for Hopper and Blackwell architectures. Data center developers can use NVIDIA optimized kernels through the FlashInfer LLM serving kernel library.

vLLM recommends using uv for Python dependency management. You can use vLLM to spin up an OpenAL-compatible web server. The following command will automatically download the model and start the server. Refer to the documentation and vLLM Cookbook guide for more details.

uv run --with vllm vllm serve openai/gpt-oss-20b

Deploy using TensorRT-LLM The optimizations are available on the NVIDIA/TensorRT-LLM GitHub repository, where developers can use the deployment guide to launch their high-performance server. The guide downloads the model checkpoints from Hugging Face. NVIDIA collaborated on the developer experience using the Transformers library with the new models. The guide then provides a Docker container and guidance on how to configure performance for both low-latency and max-throughput cases.

More than a million tokens per second with GB200 NVL72 NVIDIA engineers partnered closely with OpenAI to ensure that the new gpt-oss-120b and gpt-oss-20b models deliver accelerated performance on Day 0 across both the NVIDIA Blackwell and NVIDIA Hopper platforms.

At launch, based on early performance measurements, a single GB200 NVL72 rack-scale system is expected to serve the larger, more computationally demanding gpt-oss-120b model at 1.5 million tokens per second, or about 50,000 concurrent users. Blackwell features many architectural capabilities that accelerate inference performance. These include a second-generation Transformer Engine with FP4 Tensor Cores and fifth-generation NVIDIA NVLink and NVIDIA NVLink Switch, for high bandwidth, enabling 72 Blackwell GPUs to act as a single, massive GPU.

The performance, versatility, and pace of innovation of the NVIDIA platform enable the ecosystem to serve the latest models on Day 0 with high throughput and low cost per token.

Try the optimized model with NVIDIA Launchable Deploying with TensorRT-LLM is also available using the Python API in a JupyterLab notebook on the Open AI Cookbook as an NVIDIA Launchable directly in the build platform where developers can test out GPUs from multiple cloud platforms. You can deploy the optimized model with a single click in a pre-configured environment.

data-src=https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-png.webp alt=The image shows the console at brev.dev for users to select which type of GPU option in the Select your Compute' page, the user can select between boxes in a row of H200, H100, A100, L40s, A10 and A100 shown. class=lazyload wp-image-104187 data-srcset=https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-png.webp 1324w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-300x113-png.webp 300w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-625x236-png.webp 625w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-179x68-png.webp 179w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-768x290-png.webp 768w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-645x244-png.webp 645w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-500x189-png.webp 500w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-160x60-png.webp 160w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-362x137-png.webp 362w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/08/Brev-291x110-png.webp 291w, https://developer-blogs.nvidia.com/wp-content/uploads/202

LINK:	https://developer.nvidia.com/blog/delivering-1-5-m-tps-inference-on-nv...
	See more stories from nvidia

Most recent headlines