Chiosa
Machine Learning

Compact configurations improve hardware generation in one SECDED task

2 min
Updated
#cost-performance#compact-configuration#correctness-gating

In plain words

The paper introduces a way to make language models generate hardware designs. Instead of asking a model to write every line of a long hardware description, it can output a short list of settings, which a deterministic program expands into the design.

For one error-correcting memory-wrapper task, a model using the short configuration produced more designs that matched the reference than the same model writing code directly. The result applies to this bounded task, not hardware generation in general.

The research

For a 512-bit error-correcting memory, flat Verilog takes about 44,237 tokens, while an eight-field configuration takes 69. Can a model specify the design briefly and let a deterministic generator produce the full hardware description?

The method has a language model emit a short hardware configuration for a deterministic generator, then tests whether that representation improves correctness and feasibility.

The authors test whether changing what a model outputs can help with a task where a full hardware description is extremely long. Their method has the model provide a short configuration, which a deterministic generator expands into RTL.

With the same few-shot model and recipe, the configuration matched the reference more often than three code-like forms. This supports a representation advantage on the tested SECDED task, where correctness could be checked against a reference generator.

The evidence comes from one SECDED memory-wrapper task and generator. The small model that used the least GPU energy per correct design also silently changed 41% of tested illegal specifications into legal ones, and the out-of-distribution results had one run per regime.

Analysis

Deep

Original paperLearn the Interface, Not the Implementation: Output Representation Dominates Training Regime in LLM-Based RTL GenerationOpen original
Submitted to arXiv
Related paperMachine LearningParallel language models coordinate several token predictions at once#cost-performance#parallel-decodingRelated paperMachine LearningMatching training optimizers adds little consistent benefit#cost-performance#matrix-aware-optimizationRelated paperMachine LearningA revised test narrows the compute gap for diffusion models#cost-performance#joint-scalingRelated paperMachine LearningTwo fully trained loops can match large models with fewer parameters#cost-performance#elastic-depth#joint-scaling
Next paperParallel language models coordinate several token predictions at once