From 20e1e3a787ca2b76a3383378f6434770a9457655 Mon Sep 17 00:00:00 2001 From: michalm Date: Wed, 30 Sep 2026 08:58:49 +0200 Subject: [PATCH] dlrmv4: replace the GBS 8192 RCP with a GB300 rerun of the reference The AMD set for GBS 8192 came from three campaigns; the 3 h campaign's stalled runs were replaced by new seeds, so the 20 runs carry one run above 80 M samples while 20-seed reruns of the reference code show five. The new entry is 20 seeds chosen as ventile midpoints of a 179-run distribution and run with the unmodified reference (mlcommons/training b5e8e2a) on 2x4 GB300, uncapped, each run to the 0.75 AUC target. Olympic mean 73.02 M (was 69.07 M), std 6.69 M (was 3.81 M), checker floor 68.16 M (was 66.30 M). Other batch sizes unchanged. --- .../training_6.1.0/rcps_dlrmv4.json | 30 +++++++++---------- 1 file changed, 15 insertions(+), 15 deletions(-) diff --git a/mlperf_logging/rcp_checker/training_6.1.0/rcps_dlrmv4.json b/mlperf_logging/rcp_checker/training_6.1.0/rcps_dlrmv4.json index 0e653ef..3194104 100644 --- a/mlperf_logging/rcp_checker/training_6.1.0/rcps_dlrmv4.json +++ b/mlperf_logging/rcp_checker/training_6.1.0/rcps_dlrmv4.json @@ -1,9 +1,9 @@ { "dlrmv4_ref_8192": { "Benchmark": "dlrmv4", - "Creator": "AMD", - "When": "Reference RCPs before 6.1 submission", - "Platform": "1x8xMI350", + "Creator": "NVIDIA", + "When": "Reference code (mlcommons/training b5e8e2a) rerun on GB300 after the 6.1 RCP review, 2026-09-24", + "Platform": "2x4xGB300", "Precision": "BF16", "BS": 8192, "Hyperparams": { @@ -13,25 +13,25 @@ }, "Epochs to converge": [ 66519040, - 68812800, 66519040, 66519040, - 71106560, - 80281600, - 61931520, + 66519040, 68812800, 68812800, + 68812800, + 68812800, + 68812800, + 71106560, + 71106560, 73400320, - 66519040, 73400320, - 66519040, - 66519040, - 71106560, 73400320, - 66519040, - 71106560, - 68812800, - 68812800 + 75694080, + 80281600, + 82575360, + 84869120, + 84869120, + 87162880 ] }, "dlrmv4_ref_16384": {