Altifigence Academy
← My learning

RTL Design: CNN Accelerators

Learning outline

Learning outline
01The inference contract
Inference and trainingConcept
Convolution and cross-correlationConcept
Shapes, stride, padding and dilationConcept
Check: shape and MAC countCheck
02The numerical contract
Fixed-point bits and unitsConcept
Scale and zero pointConcept
Accumulator width, rounding and saturationConcept
Lab: execute the quantization contractLab
03Reference operators
Direct convolution referenceConcept
im2col and matrix multiplicationConcept
Connecting ReLU and max poolingConcept
Check: output processing orderCheck
04Streaming window generation
Line buffers and 3×3 windowsConcept
Ready/valid and stalled stateConcept
Lab: windows that survive pausesLab
Check: first-window timingCheck
05Mapping to PEs and memory
PEs, dataflow and MAC reuseConcept
Tiles, halos and partial sumsConcept
Bandwidth and compute boundsConcept
Schedules, tails and double bufferingConcept
06Verification and a measurable capstone
Golden models and failure classificationConcept
Lab: challenge the corner casesLab
Capstone: a small inference path and reportLab
Final check: claims and evidenceCheck

Your choice applies to this browser. Change it any time using the footer.

Necessary storage

Supports secure sign-in, requested learning records and display settings, and your privacy choices.

Always on

No advertising or analytics trackers are used. Your choice is stored for 180 days and is separate from sign-in or accepting terms.