Interactive demo of per-layer embedding costs: step one token through eight decoder layers while an arithmetic meter fills and a fetch counter adds one tiny row per layer, then use a slider to reallocate the 5B parameters between matmuls and lookup tables and watch compute, fast memory and storage change.
One token's bill, layer by layer
Arithmetic done0%
Rows fetched0
Press advance: every layer multiplies with all of its weights, and fetches one small row.
8 layers drawn; the real stack is deeper. The ratio is the point.
The allocation dial
Share of the 5B living in PLE tables: 54%
plain dense modelthe real E2B split
Per-token compute2.3B
Fast memory needed2.3B
Ordinary storage (tables)2.7B
Runs like a 2.3B model. Storage stays 5B total; only the matmul share costs compute and fast memory.