ubenchX 微基准测试（Beta）
底层 GPU 微基准测试，测量基础硬件特性。
HBM 带宽
在 NVIDIA 和 AMD GPU 上测量不同消息大小的显存拷贝带宽微基准测试。
显存带宽利用率（MBU）vs 消息大小
MBU 基于各 GPU 自身在 GPU_SPECS 中的峰值 HBM 带宽计算。
Shift+滚轮缩放 · 拖动平移 · 双击重置 · 点击数据点固定提示框
测试方法
使用 triton.testing.do_bench 计时 b.copy_(a)（float32 张量）；带宽 = 2 * bytes / time（读 + 写）。消息大小为 2 的幂次，从 8 B 到 16 GiB。
B300 SXM: NVIDIA B300 SXM6 AC | Driver: 580.159.03 | PyTorch: 2.14.0+cu130 | Triton: 3.8.0 | Container: pytorch/pytorch:2.14.0-cuda13.0-cudnn9-runtime | Peak: 8 TB/s (GPU_SPECS)
GB200 NVL72: NVIDIA GB200 | Driver: 580.126.20 | PyTorch: 2.13.0+cu130 | Triton: 3.7.1 | Container: lmsysorg/sglang:nightly-dev-cu13-20260922-582389ce (arm64) | Peak: 8 TB/s (GPU_SPECS)
B200 SXM: NVIDIA B200 | Driver: 580.159.03 | PyTorch: 2.13.0+cu130 | Triton: 3.7.1 | Container: lmsysorg/sglang nightly 2026-09-22 (582389ce) | Peak: 8 TB/s (GPU_SPECS)
H200 SXM: NVIDIA H200 | Driver: 580.173.02 | PyTorch: 2.13.0+cu130 | Triton: 3.7.1 | Container: lmsysorg/sglang:nightly-dev-cu13-20260923-06008c17 | Peak: 4.8 TB/s (GPU_SPECS)
H100 SXM: NVIDIA H100 80GB HBM3 (SXM) | Driver: 580.159.03 | PyTorch: 2.5.1+cu124 | Triton: 3.1.0 | Container: pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime | Peak: 3.35 TB/s (GPU_SPECS)
MI355X: AMD Instinct MI355X | Driver: amdgpu 6.16.6 | PyTorch: 2.11.0+rocm7.2 | Triton: 3.7.0 | Container: sglang_local-rocm724-mi35x.sqsh | Peak: 8 TB/s (GPU_SPECS)
MI325X: AMD Instinct MI325X | Driver: amdgpu 6.16.13 | PyTorch: 2.10.0+rocm7.2.4 | Triton: 3.6.0 | Container: rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0 | Peak: 6 TB/s (GPU_SPECS)
MI300X: AMD Instinct MI300X | Driver: amdgpu 6.16.13 | PyTorch: 2.10.0+rocm7.2.4 | Triton: 3.6.0 | Container: rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0 | Peak: 5.3 TB/s (GPU_SPECS)
数据来源: https://github.com/SemiAnalysisAI/InferenceX/pull/3879
数据来源: https://github.com/SemiAnalysisAI/InferenceX/pull/3874
数据来源: https://github.com/SemiAnalysisAI/InferenceX/pull/3880