OCTOPAI ENGINEERING ARTS · INTERNAL · TWO REJECTED CLAIMS
128 GB decides what fits
One box, one unified memory pool: a 122-billion-parameter model fits only because the memory is shared — and memory bandwidth decides how many people it serves.
16 interactive chat streams15.8 tok/s each
35B-A3B, one user21 tok/s · hard-tier code 88%
122B GPTQ, one user10 tok/s · best hard-tier chat 78%
vs NVIDIA GB101.3× one user · level at 16
Grade rule≥10 tok/s AND passes the quality gate
H1 · MEASURED · OFFICIAL ROCM/VLLM STACK
H2 · WITHIN NOISE OF GB10
H3 · FACT · A RYZEN AI HALO BOX, NOT A DELL SKU
SOURCE · HALO STUDY · METRUM AI BENCH-CLI · OFFICIAL ROCM/VLLM STACK
OCTOPAI ENGINEERING ARTS · INTERNAL · SOFTWARE MOVED THE NUMBERS
HARD-TIER CODE · SAME BOX · SAME MODEL · ONE FLAG CHANGED
BASELINE
29 / 40
WITH N-GRAM SPECULATION
6 / 40 · 25 NEW SYNTAX ERRORS
A SPEED-ONLY BENCHMARK WOULD HAVE SCORED THIS AS AN IMPROVEMENT.
THE QUALITY GATE CAUGHT IT.
A2 · PRIMARY · OCCUPANCY > 80%
A3 · SINGLE WAVES · NOT SUSTAINED
A1 · SQLITE AGENTS · PINNED EPYC 9R14 · AWS
SOURCE · AGENTICBENCH CORE SIZING PILOT #cpuonly #results #holdout · CURRENT CAPTURE
ALL-SMI TELEMETRY · 47-METRIC GAP MATRIX · BEFORE → AFTER
CORRECT
11 → 17
PARTIAL
8 → 16
WRONG
2 → 0
MISSING
26 → 14
GPU POWER ON THIS APU WAS REPORTING SOCKET POWER · NOW 0.104 W GFX vs 114.6 W SOCKET UNDER CPU-ONLY LOAD
B5 · MEASURED · ALL-SMI FORK · 47 METRICS
CROSS-CHECKED vs nvidia-smi AND amd-smi
SOURCE · HALO STUDY · TELEMETRY VALIDATION · B5