From 6c6464a31a14ec70adc1ea32c2acf5f34d60b23d Mon Sep 17 00:00:00 2001
From: Frank <frank@dankware>
Date: Fri, 18 Sep 2026 00:28:29 -0700
Subject: [PATCH] frank-medium: record 2026-09-18 rebalance measurement, keep
 35:30

Isolated-port (:7790) 2k-token decode with AVG rocm-smi util over the
request (peak util lies: a layer split touches both cards every token):

  35:30 draft ROCm1  31.52 t/s  util0 41%  util1 98%  <- deployed
  36:29 draft ROCm1  31.49 t/s  util0 43%  util1 98%
  35:30 draft ROCm0  31.52 t/s  util0 42%  util1 98%
  36:29 draft ROCm0  31.48 t/s  util0 43%  util1 98%

The x1 card (ROCm1) stays the 98% bottleneck in every configuration.
Moving the MTP drafter to the x16 card does not flip the balance: the
per-step cost is the verify phase (both cards run the full target) plus
the logits, which are pinned to the card holding the last layers. A
target-layer move (36:29) buys 2 points on the x16 card and erodes its
VRAM margin (fit free 0.45 GiB) for no throughput change. The x1 card is
compute-bound, not PCIe-bound (~5.2 MB activation per ubatch on the link),
so saturating it costs no tokens. 35:30 is the balance; no split change.
---
 scripts/frank-medium.env | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/scripts/frank-medium.env b/scripts/frank-medium.env
index 6a38573..62d88c3 100644
--- a/scripts/frank-medium.env
+++ b/scripts/frank-medium.env
@@ -164,6 +164,26 @@ SPLIT_MODE=layer
 # 1.483 GiB slack) -- both gates fail.  35:30 (fit: 30.59 / 30.01, gap
 # 0.58) is the farthest safe shift toward the healthy card at this pool.
 # Do not move it without re-measuring.
+
+# MEASURED 2026-09-18 (isolated :7790, 2k-token decode, AVG rocm-smi util over
+# the request -- peak util lies, a layer split touches both cards every token):
+#
+#   split   draft_dev  decode   AVG util0(x16)  AVG util1(x1)
+#   35,30   ROCm1      31.52    41 %            98 %           <- deployed
+#   36,29   ROCm1      31.49    43 %            98 %
+#   35,30   ROCm0      31.52    42 %            98 %
+#   36,29   ROCm0      31.48    43 %            98 %
+#
+# The x1 card stays the 98 % bottleneck in EVERY configuration.  The lever that
+# was tried is moving the MTP drafter to the x16 card (SPEC_DRAFT_DEVICE=ROCm0):
+# it does NOT flip the balance, because the drafter is a small nextn head and the
+# per-step cost is dominated by the VERIFY phase -- both cards run the full target
+# -- plus the LOGITS, which are pinned to the card holding the last layers.  A
+# target-layer move (36:29) buys 2 points on the x16 card and erodes its VRAM
+# margin (fit: free 0.45 GiB) for nothing.  The x1 card is COMPUTE-bound, not
+# PCIe-bound (~5.2 MB of activation per 512-token ubatch crosses the link);
+# saturating it costs no tokens beyond what any R9700 would.  35:30 is the
+# balance; do not chase the x16 card past it.
 TENSOR_SPLIT=35,30
 
 NGL=999