[Klaud Cold] feat(ubenchX): add B300, B200, H200, GB200 mem_bw results / [Klaud Cold] feat(ubenchX)：新增 B300、B200、H200、GB200 mem_bw 测试结果

[functionstackx]

Add ubenchX/mem_bw results for B300, B200, H200, and GB200, next to the existing H100 SXM file. 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	Driver	Container	torch / Triton	Peak (16 GiB)
b300_2026-10-09.csv	NVIDIA B300 SXM6 AC	580.159.03	pytorch/pytorch:2.14.0-cuda13.0-cudnn9-runtime	2.14.0+cu130 / 3.8.0	6,656 GB/s
b200_2026-10-09.csv	NVIDIA B200	580.159.03	cached lmsysorg/sglang nightly 2026-09-22 (582389ce)	2.13.0+cu130 / 3.7.1	6,605 GB/s
h200_2026-10-09.csv	NVIDIA H200	580.173.02	cached lmsysorg/sglang:nightly-dev-cu13-20260923-06008c17	2.13.0+cu130 / 3.7.1	4,305 GB/s
gb200_2026-10-09.csv	NVIDIA GB200	580.126.20	cached lmsysorg/sglang:nightly-dev-cu13-20260922-582389ce (arm64)	2.13.0+cu130 / 3.7.1	6,892 GB/s

Clusters: B300 dsxe-sa-b300-prd0, B200 nsc-sa (im-b200), H200 login-h200-1, GB200 watchtower (batch-navy). GB300 was not run: its login node (login-gb300) is offline on Tailscale.

Caveats: the container and torch versions differ across GPUs (the H100 run used torch 2.5.1). Copy kernels this simple should not be sensitive to that, but the runs are not strictly like for like. All runs are single samples on a single GPU.

中文

在现有 H100 SXM 结果旁新增 B300、B200、H200 和 GB200 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	驱动	容器	torch / Triton	峰值（16 GiB）
b300_2026-10-09.csv	NVIDIA B300 SXM6 AC	580.159.03	pytorch/pytorch:2.14.0-cuda13.0-cudnn9-runtime	2.14.0+cu130 / 3.8.0	6,656 GB/s
b200_2026-10-09.csv	NVIDIA B200	580.159.03	已缓存的 lmsysorg/sglang nightly 2026-09-22（582389ce）	2.13.0+cu130 / 3.7.1	6,605 GB/s
h200_2026-10-09.csv	NVIDIA H200	580.173.02	已缓存的 lmsysorg/sglang:nightly-dev-cu13-20260923-06008c17	2.13.0+cu130 / 3.7.1	4,305 GB/s
gb200_2026-10-09.csv	NVIDIA GB200	580.126.20	已缓存的 lmsysorg/sglang:nightly-dev-cu13-20260922-582389ce（arm64）	2.13.0+cu130 / 3.7.1	6,892 GB/s

集群：B300 为 dsxe-sa-b300-prd0，B200 为 nsc-sa（im-b200），H200 为 login-h200-1，GB200 为 watchtower（batch-navy）。GB300 未运行：其登录节点（login-gb300）在 Tailscale 上处于离线状态。

注意事项：各 GPU 使用的容器和 torch 版本不同（H100 使用 torch 2.5.1）。这类简单的拷贝 kernel 对此应不敏感，但各次运行并非严格对等。所有结果均为单块 GPU 上的单次测量。

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]

LGTM, straightforward data-only addition.

What was reviewed: the full diff (4 new CSV files, 33 lines each, no code/config changes) was read in full; verified each file has the same header/row structure (bytes,time_ms,bandwidth_gbps) and a monotonically increasing power-of-two byte sweep from 8 B to 16 GiB. Spot-checked peak bandwidth numbers against known HBM specs for each platform (H200 ~4.3 TB/s, B200/B300 ~6.6 TB/s, GB200 ~6.9 TB/s) and they fall in plausible achieved-vs-theoretical ranges. No source, docs, or config files are touched, so no CODEOWNERS or perf-changelog scope applies.

Extended reasoning...

The diff adds four new, data-only CSV result files under ubenchX/mem_bw/results/ (132 lines total insertions, no deletions, no other files touched) recording a memory-bandwidth sweep for four GPU platforms produced by a pre-existing benchmark script. There is no code, config, or security-sensitive surface involved, the file shape is consistent across all four CSVs, and the reported peak bandwidth figures are plausible for each GPU's HBM generation, so this is a simple, mechanical data addition a human does not need to block on.