NeuroEtch
Loading
ICsprout55 in_proj dual-MAC 32-row
Throughput scaling on the in_proj_qkv chain: same 32 rows as CF-4, half the cycles via dual bank fetch. STA closes at 65 MHz.
Key points
Tensor
in_proj_qkv L0
Same 32 rows as CF-4; dual-port fetch.
MAC
32 lanes
64→32 cycles per dot versus CF-4.
Cells
~71k mapped
158,887 instances after tap/CTS/fill.
DRT
0 violations
500×500 µm die at density 0.50.
Timing (exp.)
65 MHz
Placement-RC sweep on repaired DEF.
CF-10 mirrors CF-9 on the in_proj tensor chain: thirty-two-row ROM scaling plus dual-MAC throughput, with machine-checkable mapped and physical proofs.
Still thirty-two rows of one projection tensor — not full-model on chip.