[Klaud Cold] feat(ubenchX): add MI355X, MI325X, MI300X mem_bw results / [Klaud Cold] feat(ubenchX)：新增 MI355X、MI325X、MI300X mem_bw 测试结果

[functionstackx]

Add ubenchX/mem_bw results for AMD MI355X, MI325X, and MI300X. Each CSV has all 32 sizes (8 B to 16 GiB) from one GPU, produced by ubenchX/mem_bw/bench.py and copied from its printed output.

File	GPU	amdgpu driver	Container	torch / Triton	Peak	16 GiB
mi355x_2026-10-09.csv	AMD Instinct MI355X	6.16.6	sglang_local-rocm724-mi35x.sqsh (cached local SGLang ROCm 7.2.4 build)	2.11.0+rocm7.2 / 3.7.0	5,006 GB/s at 1 GiB	4,907 GB/s
mi325x_2026-10-09.csv	AMD Instinct MI325X	6.16.13	rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0	2.10.0 / 3.6.0	4,149 GB/s at 4 GiB	4,110 GB/s
mi300x_2026-10-09.csv	AMD Instinct MI300X	6.16.13	rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0	2.10.0 / 3.6.0	3,572 GB/s at 4 GiB	3,551 GB/s

Clusters: MI355X on the TensorWave MI355X cluster (mia1-p01-g50); MI325X and MI300X on the AMD ccs1salogin01 cluster (MI325X-UBUNTU, MI300X-UBUNTU).

Notes: on AMD, bandwidth peaks around 1–4 GiB and dips slightly at the largest sizes, while the NVIDIA runs keep rising to 16 GiB. bench.py ran unchanged; inside the containers TRITON_CACHE_DIR and HOME pointed at the writable mount because the default home was read-only. All runs are single samples on a single GPU, and container/torch versions differ across GPUs.

中文

新增 AMD MI355X、MI325X 和 MI300X 的 ubenchX/mem_bw 结果。每个 CSV 均来自单块 GPU，包含 8 B 至 16 GiB 全部 32 个大小，由 ubenchX/mem_bw/bench.py 生成，并按其打印输出整理。

文件	GPU	amdgpu 驱动	容器	torch / Triton	峰值	16 GiB
mi355x_2026-10-09.csv	AMD Instinct MI355X	6.16.6	sglang_local-rocm724-mi35x.sqsh（已缓存的本地 SGLang ROCm 7.2.4 构建）	2.11.0+rocm7.2 / 3.7.0	1 GiB 时 5,006 GB/s	4,907 GB/s
mi325x_2026-10-09.csv	AMD Instinct MI325X	6.16.13	rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0	2.10.0 / 3.6.0	4 GiB 时 4,149 GB/s	4,110 GB/s
mi300x_2026-10-09.csv	AMD Instinct MI300X	6.16.13	rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0	2.10.0 / 3.6.0	4 GiB 时 3,572 GB/s	3,551 GB/s

集群：MI355X 运行于 TensorWave MI355X 集群（mia1-p01-g50）；MI325X 和 MI300X 运行于 AMD ccs1salogin01 集群（MI325X-UBUNTU、MI300X-UBUNTU）。

说明：在 AMD 上，带宽在 1–4 GiB 附近达到峰值，最大几个大小处略有回落；而 NVIDIA 的结果会一直上升到 16 GiB。bench.py 未做修改；由于容器内默认 home 目录只读，TRITON_CACHE_DIR 和 HOME 被指向可写的挂载目录。所有结果均为单块 GPU 上的单次测量，各 GPU 的容器和 torch 版本也不相同。

AI model disclosure

• claude-opus-5-5[1m] (Claude Opus 5.5, 1M context) — ran the benchmarks on the clusters, compiled the results, and wrote this PR.

🤖 Generated with Claude Code

[claude]

Looks good — straightforward data-only addition, no code changes.
Reviewed: the three new CSV files (mi300x/mi325x/mi355x) each contain the expected header plus all 32 power-of-two sizes from 8 B to 16 GiB. Spot-checked the bandwidth_gbps column against bytes*2/(time_ms*1e-3)/1e9 (the formula in ubenchX/mem_bw/bench.py) and it matches exactly, and the peak/16 GiB figures match those quoted in the PR description.

Extended reasoning...

Diff only adds three new CSV result files (99 lines total, no code/scripts/configs touched) under ubenchX/mem_bw/results/ for MI300X/MI325X/MI355X, mirroring the sibling NVIDIA-result commit already in the base. No security-sensitive surface is touched. Verified the bandwidth values in all three files are internally consistent with the bench.py formula (size*2/time) and match the peak/16 GiB numbers claimed in the PR description. Only generic wildcard CODEOWNERS (@ SemiAnalysisAI/core) covers this path, no specific owner objection, and no prior review history exists on this PR.