Interactive demo of per-layer embeddings: a matrix multiply uses every weight per token while an embedding table uses one row, and a toggle compares E2B's compute and memory cost with an imaginary plain 5B model.
One token's cost: multiply vs look up
Multiply versus lookup
Matrix multiply
Every weight works
Embedding table
One row per token
touched for this token
idle
Where the 5 billion live
E2B with PLE
Imaginary plain 5B
Memory placement
Fast chip memory
Scarce and quick
Ordinary memory
Big and slower
Core transformer
2.3B of matmuls
PLE tables, 2.7B
Fetched row by row
Compute per token
2.3B
Fast memory needed
2.3B
Ordinary memory used
2.7B