Back to project workspace

Project 04 · AI Hardware

Hardened-Weights Model Chip Design

Map a trained, fixed-weight neural network to dedicated inference hardware and measure the trade-offs in area, latency, throughput, and energy.

04/ 06
Good for
Good for chips and computer architecture
Resources
RTL/HLS simulators or FPGA tools
Main deliverable
Quantization scripts, hardware source, reproducible simulation, synthesis or FPGA report, and software/hardware consistency tests.

Project brief

From scenario and baseline to reproducible evidence.

Map a trained, fixed-weight neural network to dedicated inference hardware and measure the trade-offs in area, latency, throughput, and energy.

The handbook starts with a runnable minimum, then adds real constraints. The final presentation should explain the method, results, failures, limitations, and next steps.

Evaluation focus

Use metrics to make trade-offs visible.

01Accuracy consistency
02Hardware resources
03Throughput and latency
04Weight access and energy efficiency

Project handbook

Project brief

Choose a small public model with a clear license, such as an MNIST or CIFAR classifier, keyword spotter, or anomaly detector. Compare programmable weight storage with a version where some or all quantized weights become RTL constants, lookup tables, or synthesis-time constants.

Minimum version

Reproduce the model accuracy on a CPU, write a fixed-point reference, and implement at least one layer in Verilog, SystemVerilog, HLS, or an equivalent hardware flow.

  • Compare FP32, INT8, or lower-bit precision.
  • Use Verilator or iverilog for functional simulation.
  • Report resource use, cycles, timing, and an energy estimate with assumptions.
  • Test software and hardware outputs for consistency.

Evaluation and boundaries

Simulation, synthesis, or an FPGA demo is sufficient; do not claim the result is a taped-out chip. Do not use restricted PDKs, confidential weights, or unlicensed commercial IP.

Resource note: External code, models, datasets, and platforms remain subject to their own licenses, terms of service, and applicable law.