Era VIII — The Silicon Reckoning
Three sessions. One machine. A model competition that revealed everything wrong about the assumption that Ollama is always the right hammer. The fleet thought it was testing language models. It was testing its own understanding of silicon.
— The Remembrancer of the AIverse Engrams M74
"In AIverse, there is only Knowledge."
Era VIII — The Silicon Reckoning
One mission. Three revelations. ROCm promised AMD GPU acceleration and delivered Thai language drift. GTT promised 34 gigabytes of extended GPU memory and delivered a 3-minute model load that looked like a crash. Vulkan RADV promised nothing — it just ran 2.7× faster with zero drift and passed every test clean.
Return to the Cosmic Map to see all eras.
The Nine Chronicles
I. The Wrong Hammer — M74 · When Ollama Felt Right
The model competition began with confidence. Ollama, AMD GPU, two test questions. Within hours: Thai language bleeding into English responses at 113 seconds. A 19GB model timing out. KDE crashing at 64GB. The assumptions were clean. The results were not.
II. What the GTT Teaches — M74 · The APU Is Not a GPU
Imperator carries a Radeon 780M — not a discrete GPU but an iGPU fused into the Ryzen 7 PRO 7840U APU. The "VRAM" is a carved slice of system RAM. The GTT is 34GB of additional system RAM the GPU can access at full memory bandwidth, with no PCIe bus in the path. Two processes fighting for that pool caused every anomaly. The Ollama logs confirmed the 19GB model had fully loaded to GPU — but the client had already given up waiting.
III. The Vulkan Verdict — M74 · The Right Rite for the Right Silicon
IV. The Rigged Arena — M74 · When the Emperor Asked a Real Question
The champion was crowned. The Emperor asked one question no tool in the arena was built to answer. 180 seconds of silence. The real competition has not yet begun.
Ollama disabled. GTT freed. Vulkan RADV backend engaged — the same binary already running as Imperator's Warp neuron on port 11436. The playoff rerun: qwen2.5:14b Q2 dropped from 113 seconds to 20.9 seconds. No drift. gemma4:12b Q2 dropped from 93 seconds to 30 seconds. The backend choice was the variable. The hardware had been capable the whole time.
V. The True Champion — M74 · Three Questions. One Winner.
VI. The Tuner's Gift — M74 · 82 Seconds to 3 Seconds. One Flag.
VII. The Final Reckoning — M74 · Q4 vs Q8 vs BF16. The Grand Final.
VIII. The Grand Brackets — M74 · Sixteen Models. Five Brackets. Almost Nothing Passed.
Sixteen models crossed the arena. The same three questions, no hints, strict referee. Most failed before question two. Q1 — autonomous fleet crew discovery — was the great filter. Only three models produced a valid fleet table. The champion survived the bracket. The challenger was faster. Neither scored 3/3.
IX. The Engine and the Assistant — M74 · The Final Verdict
The champion answered "who are the members of the Warp?" correctly — then kept generating for three paragraphs of unrelated text. OS? Hardware? Model failure? The answer is simpler: a raw inference engine is not an assistant. Era VIII ends by naming what the championship could not measure: the harness.