From c2edb273ca8be3ba599ce74a0e530ebd579a693a Mon Sep 17 00:00:00 2001
From: jgrusewski <jgrusewski@localhost>
Date: Sun, 19 Apr 2026 09:01:34 +0200
Subject: [PATCH] =?UTF-8?q?perf:=20batch=5Fsize=2016384=20=E2=86=92=208192?=
 =?UTF-8?q?=20=E2=80=94=202x=20faster=20GEMMs,=20more=20frequent=20updates?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

At 16384, each GEMM saturates 97% of 132 SMs — no room for multi-stream
parallelism. At 8192, GEMMs use ~50% SMs, allowing branch streams and
aux parallelism to fill idle SMs. 2x more steps/epoch but each ~2x faster.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 config/gpu/h100.toml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/config/gpu/h100.toml b/config/gpu/h100.toml
index 0c63ad7bb..9ca38f725 100644
--- a/config/gpu/h100.toml
+++ b/config/gpu/h100.toml
@@ -1,6 +1,6 @@
 # H100 PCIe/SXM (80GB VRAM) -- full production
 [training]
-batch_size = 16384
+batch_size = 8192
 num_atoms = 52
 buffer_size = 500000
 hidden_dim_base = 256