Measure the workload before freezing the machine.
ProtonSim is a workload and architecture modeling direction for connecting transformer behavior to compute, memory, and dataflow decisions.
Validated public workload results are not yet available
No benchmark result is rendered until a workload is validated.
RTL is an expensive place to discover the wrong architecture.
A workload model can expose operator mix, data movement, arithmetic intensity, utilization limits, and prefill/decode differences before implementation choices harden.
Define what the system is expected to execute.
Model family, parameter envelope, batch, context, quantization, operator coverage, and execution phase form the first traceable input.
Break the model into work and movement.
GEMM, GEMV, reductions, normalization, positional operations, elementwise work, weight reads, activation traffic, and KV-cache traffic are characterized separately.
Compute throughput matters only when data can arrive.
Roofline-style analysis is intended to compare arithmetic demand with local and external bandwidth limits without presenting unvalidated performance results.
Sweep the decisions that materially change the machine.
Candidate sweeps include memory bandwidth, scratchpad capacity, vector width, vector lanes, matrix dimensions, dataflows, and prefill/decode policy.
Every recommendation should point back to workload evidence.
The intended output is a versioned architecture report that links assumptions, model inputs, analytical results, and the decision they support.
Illustrative controls are not benchmark results.
The current website demonstrates the intended design space only. Accuracy, cycle correlation, supported models, and performance remain to be validated before publication.