Skip to content

Update H100 and MI300X distributed benchmark data #391

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
inaniloquentee merged 1 commit into from
Sep 8, 2026
+682 −336
Merged

Update H100 and MI300X distributed benchmark data #391

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
59 changes: 42 additions & 17 deletions benchmarks/results/pr325_rocm_mi300x/cuda_cpu_comparison.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,32 @@
"pull_request": "https://github.com/RL-Align/RL-Kernel/pull/321",
"cuda_commit": "8576fa4bf449734ae99e9b50be8756bb282a8916",
"h100_triton_replay_commit": "e64abab904880b877d26d04c0cfad020b992aa51",
"note": "User-supplied measurements; cross-hardware values are context only and are not hardware-normalized."
"note": "User-supplied measurements; cross-hardware values are context only and are not hardware-normalized.",
"official_h100_update": {
"source": "User-supplied H100 table on 2026-09-06",
"topologies": ["tp2", "tp2_sp", "tp4", "tp2_cp2", "tp2_cp2_sp", "tp8", "tp4_cp2", "tp4_cp2_sp"],
"directions": ["forward", "train_fwd_bwd"],
"fields": ["official_h100_ms", "official_h100_summary_ms"],
"git_commit": "98dcd38fb635e1e0eab9035aa4c4b40483cc36b8",
"gpu": "NVIDIA H100 80GB",
"torch": "2.13.0+cu130",
"cuda": "13.0",
"units": "ms",
"shape": {"tokens": 32, "hidden": 4096, "intermediate": 12288},
"dtype": "bfloat16",
"warmup": 5,
"samples": 20,
"training_samples": 10,
"timing": "synchronized wall clock; slowest rank per sample",
"baseline": "F.linear + NCCL at the same TP/CP/SP topology; TP=1 is used only for untimed RL-Kernel bitwise checks",
"validation": {"reference": "RL-Kernel CUDA TP=1", "output_mismatch": 0, "hidden_gradient_mismatch": 0, "weight_gradient_mismatch": 0, "repeat_mismatch": 0, "train_infer_mismatch": 0},
"environment_variables": {"NCCL_NVLS_ENABLE": "0", "NCCL_IB_DISABLE": "1", "RL_KERNEL_DET_GEMM_BACKEND": "sm90"},
"commands": [
"env PYTHONPATH=/tmp/rlk-head-98dcd38-build-GYkK6iE5/repo CUDA_VISIBLE_DEVICES=0,1 NCCL_NVLS_ENABLE=0 NCCL_IB_DISABLE=1 RL_KERNEL_DET_GEMM_BACKEND=sm90 FFN_WORLD_SIZE=2 FFN_WARMUP=5 FFN_SAMPLES=20 FFN_TRAINING_SAMPLES=10 /home/ellm/miniconda3/envs/hsr_env_rl_kernel/bin/python /tmp/rlk-head-98dcd38-build-GYkK6iE5/repo/run_h100_distributed_ffn_data.py",
"env PYTHONPATH=/tmp/rlk-head-98dcd38-build-GYkK6iE5/repo CUDA_VISIBLE_DEVICES=0,1,2,3 NCCL_NVLS_ENABLE=0 NCCL_IB_DISABLE=1 RL_KERNEL_DET_GEMM_BACKEND=sm90 FFN_WORLD_SIZE=4 FFN_WARMUP=5 FFN_SAMPLES=20 FFN_TRAINING_SAMPLES=10 /home/ellm/miniconda3/envs/hsr_env_rl_kernel/bin/python /tmp/rlk-head-98dcd38-build-GYkK6iE5/repo/run_h100_distributed_ffn_data.py"
],
"note": "All 16 official distributed rows are updated from user-supplied H100 measurements. CUDA and other historical rows remain unchanged; top-level environment and methodology describe those historical data. All H100 world sizes used 5/20/10 warmup/forward/training settings. ROCm uses 10/50/20 CLI settings (its distributed training halves warmup to 5); shapes, topology and timing aggregation match, but sample counts differ. The user discarded H100 10/50/20 reruns contaminated by competing GPU 0-3 workloads. The supplied H100 runner and individual samples are not available in this repository."
}
},
"environment": {
"gpu": "NVIDIA H100 80GB HBM3",
Expand Down Expand Up @@ -39,22 +64,22 @@
{"tokens": 32, "direction": "train_fwd_bwd", "official_h100_ms": 0.7842, "cuda_h100_ms": 9.1635, "triton_h100_ms": 5.4261, "official_cpu_ms": 65.9689}
],
"distributed": [
{"name": "tp2", "direction": "forward", "official_h100_ms": 0.1552, "cuda_h100_ms": 2.7896},
{"name": "tp2", "direction": "train_fwd_bwd", "official_h100_ms": 0.6301, "cuda_h100_ms": 7.8125},
{"name": "tp2_sp", "direction": "forward", "official_h100_ms": 0.1552, "cuda_h100_ms": 3.3310},
{"name": "tp2_sp", "direction": "train_fwd_bwd", "official_h100_ms": 0.6301, "cuda_h100_ms": 11.7695},
{"name": "tp4", "direction": "forward", "official_h100_ms": 0.1530, "cuda_h100_ms": 2.1687},
{"name": "tp4", "direction": "train_fwd_bwd", "official_h100_ms": 0.5714, "cuda_h100_ms": 9.3954},
{"name": "tp2_cp2", "direction": "forward", "official_h100_ms": 0.1530, "cuda_h100_ms": 2.8261},
{"name": "tp2_cp2", "direction": "train_fwd_bwd", "official_h100_ms": 0.5714, "cuda_h100_ms": 16.1722},
{"name": "tp2_cp2_sp", "direction": "forward", "official_h100_ms": 0.1530, "cuda_h100_ms": 3.5814},
{"name": "tp2_cp2_sp", "direction": "train_fwd_bwd", "official_h100_ms": 0.5714, "cuda_h100_ms": 16.7894},
{"name": "tp8", "direction": "forward", "official_h100_ms": 0.1537, "cuda_h100_ms": 2.3206},
{"name": "tp8", "direction": "train_fwd_bwd", "official_h100_ms": 0.5006, "cuda_h100_ms": 10.1375},
{"name": "tp4_cp2", "direction": "forward", "official_h100_ms": 0.1537, "cuda_h100_ms": 2.2141},
{"name": "tp4_cp2", "direction": "train_fwd_bwd", "official_h100_ms": 0.5006, "cuda_h100_ms": 16.5617},
{"name": "tp4_cp2_sp", "direction": "forward", "official_h100_ms": 0.1537, "cuda_h100_ms": 3.2972},
{"name": "tp4_cp2_sp", "direction": "train_fwd_bwd", "official_h100_ms": 0.5006, "cuda_h100_ms": 17.1457}
{"name": "tp2", "direction": "forward", "official_h100_ms": 0.1653, "cuda_h100_ms": 2.7896, "official_h100_summary_ms": {"median_ms": 0.1653, "p95_ms": 0.2055, "min_ms": 0.1420, "max_ms": 0.3013}},
{"name": "tp2", "direction": "train_fwd_bwd", "official_h100_ms": 0.6347, "cuda_h100_ms": 7.8125, "official_h100_summary_ms": {"median_ms": 0.6347, "p95_ms": 0.6955, "min_ms": 0.5826, "max_ms": 0.6997}},
{"name": "tp2_sp", "direction": "forward", "official_h100_ms": 0.1974, "cuda_h100_ms": 3.3310, "official_h100_summary_ms": {"median_ms": 0.1974, "p95_ms": 0.2161, "min_ms": 0.1807, "max_ms": 0.2273}},
{"name": "tp2_sp", "direction": "train_fwd_bwd", "official_h100_ms": 0.7162, "cuda_h100_ms": 11.7695, "official_h100_summary_ms": {"median_ms": 0.7162, "p95_ms": 1.0543, "min_ms": 0.6713, "max_ms": 1.2747}},
{"name": "tp4", "direction": "forward", "official_h100_ms": 0.1742, "cuda_h100_ms": 2.1687, "official_h100_summary_ms": {"median_ms": 0.1742, "p95_ms": 0.8594, "min_ms": 0.1545, "max_ms": 1.4339}},
{"name": "tp4", "direction": "train_fwd_bwd", "official_h100_ms": 0.6460, "cuda_h100_ms": 9.3954, "official_h100_summary_ms": {"median_ms": 0.6460, "p95_ms": 0.7029, "min_ms": 0.5959, "max_ms": 0.7232}},
{"name": "tp2_cp2", "direction": "forward", "official_h100_ms": 0.1600, "cuda_h100_ms": 2.8261, "official_h100_summary_ms": {"median_ms": 0.1600, "p95_ms": 0.8807, "min_ms": 0.1411, "max_ms": 1.4457}},
{"name": "tp2_cp2", "direction": "train_fwd_bwd", "official_h100_ms": 1.2175, "cuda_h100_ms": 16.1722, "official_h100_summary_ms": {"median_ms": 1.2175, "p95_ms": 2.2282, "min_ms": 1.1868, "max_ms": 2.2589}},
{"name": "tp2_cp2_sp", "direction": "forward", "official_h100_ms": 0.2425, "cuda_h100_ms": 3.5814, "official_h100_summary_ms": {"median_ms": 0.2425, "p95_ms": 0.2821, "min_ms": 0.2271, "max_ms": 0.3538}},
{"name": "tp2_cp2_sp", "direction": "train_fwd_bwd", "official_h100_ms": 1.5878, "cuda_h100_ms": 16.7894, "official_h100_summary_ms": {"median_ms": 1.5878, "p95_ms": 1.6838, "min_ms": 1.4295, "max_ms": 1.7189}},
{"name": "tp8", "direction": "forward", "official_h100_ms": 0.3962, "cuda_h100_ms": 2.3206, "official_h100_summary_ms": {"median_ms": 0.3962, "p95_ms": 1.1780, "min_ms": 0.3521, "max_ms": 1.2054}},
{"name": "tp8", "direction": "train_fwd_bwd", "official_h100_ms": 1.2062, "cuda_h100_ms": 10.1375, "official_h100_summary_ms": {"median_ms": 1.2062, "p95_ms": 1.4400, "min_ms": 1.1445, "max_ms": 1.4588}},
{"name": "tp4_cp2", "direction": "forward", "official_h100_ms": 0.3806, "cuda_h100_ms": 2.2141, "official_h100_summary_ms": {"median_ms": 0.3806, "p95_ms": 1.4873, "min_ms": 0.3540, "max_ms": 2.5012}},
{"name": "tp4_cp2", "direction": "train_fwd_bwd", "official_h100_ms": 1.5640, "cuda_h100_ms": 16.5617, "official_h100_summary_ms": {"median_ms": 1.5640, "p95_ms": 2.7782, "min_ms": 1.4680, "max_ms": 2.9183}},
{"name": "tp4_cp2_sp", "direction": "forward", "official_h100_ms": 0.5009, "cuda_h100_ms": 3.2972, "official_h100_summary_ms": {"median_ms": 0.5009, "p95_ms": 0.5523, "min_ms": 0.4745, "max_ms": 0.5670}},
{"name": "tp4_cp2_sp", "direction": "train_fwd_bwd", "official_h100_ms": 1.7319, "cuda_h100_ms": 17.1457, "official_h100_summary_ms": {"median_ms": 1.7319, "p95_ms": 1.9220, "min_ms": 1.7104, "max_ms": 1.9283}}
],
"dtype_accuracy": {
"tokens": 8,
Expand Down
Binary file modified benchmarks/results/pr325_rocm_mi300x/distributed_ffn_overhead.png
View file
Open in desktop
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Loading
Loading