Highlights
- Decode is treated as a separate workload from prefill.
- The primary artifact is a bytes/token ledger.
- Data release will live in an external archive record.
Artifact card
| Track | Memory systems / accelerator architecture |
|---|---|
| Status | Draft |
| Interface/spec | Pending |
| Benchmark schema | Pending |
| Known limitations | Listed in release manifest |
| Version compatibility | Pending |
| External archive | Archive pending |
Abstract
This draft separates prefill from decode and builds a first PCCX traffic ledger for weight reads, KV-cache reads, activation movement, and runtime overhead.
Motivation
This article defines a PCCX research artifact that can be cited, audited, and updated without turning draft evidence into marketing copy.
Evidence snapshot
The article records what is currently known, what remains unresolved, and what artifact must exist before stronger claims are made.
Method outline
The release package should include metadata, source logs, scripts, figures, checksums, citation files, and a plain-English limitation note.
Reproduce this result
- Open the external archive record when published.
- Download the manifest and artifact bundle.
- Run the parser or notebook listed in the manifest.
- Regenerate the figure and compare the checksum.
External archive
| Provider | Zenodo planned |
|---|---|
| Status | Archive pending |
| DOI | Pending |
| Record | Not yet published |
Code availability
Analysis scripts and raw parsers should be released through the project repository or a DOI-backed external archive, then linked here.
Commercial relevance
Defines memory-system requirements for future Evaluation, ProCore scheduler work, and ASICKit memory hierarchy planning.
References
- PCCX Research. Article archive record, version 0.1.0-draft.
- PCCX artifact metadata and citation records.