Systems Engineering Playbook: Optimizing Qwen 3.5-397B MoE on Ironwood (TPU7x)
To optimize this track, the performance team implemented a series of algorithmic fusions and precision co-designs:1. Causal Conv1D FusionThe GDN recurrent update is preceded by a causal 1D convolution (K=4). Initially, this was compiled as an independent operation, forcing the intermediate convolution outputs to be written to and read from HBM.
- ▪To optimize this track, the performance team implemented a series of algorithmic fusions and precision co-designs:1.
- ▪Causal Conv1D FusionThe GDN recurrent update is preceded by a causal 1D convolution (K=4).
- ▪Initially, this was compiled as an independent operation, forcing the intermediate convolution outputs to be written to and read from HBM.
Google Developers Blog files mainly under programming. We currently carry 20 of its stories.
Story provenance
Source · retrieval · rights · ranking — open for full record
inspect →
Story provenance
Attribution is not the same as permission. This drawer separates discovery metadata, excerpts, WeSearch-generated summaries, reuse status, and whether the publisher receives the visit. Nothing here claims a legal grant the publisher has not made.
Record
| Original publisher | Google Developers Blog |
| Canonical URL | https://developers.googleblog.com/systems-engineering-playbook-optimizing-qwen-35-397b-moe-on-ironwood-tpu7x/ |
| Publication time | Not provided by source |
| Retrieval time | 2026-07-25T23:18:53.148Z |
| Last seen | 2026-07-25T23:19:03.067Z |
| Headline source | Publisher (no WeSearch rewrite) |
| Excerpt source | publisher body |
| Excerpt method | First ~120 words (~800 chars) of extracted publisher body, fair-use limited. |
| Summary | WeSearch · cerebras-chat (WeSearch summarizer) |
| Summary source text | contentText |
| Citation coverage | Summary is a WeSearch-generated derivative; primary citation is the original publisher URL. |
| Cluster | xrwTd2_l1L0v |
| Cluster logic | Grouped by semantic title/content similarity across sources within a rolling window. Same-publisher template collisions are excluded from coverage comparison. |
| Ranking reason | Story pages are not engagement-ranked. Hub feeds use recency, with optional source-diversified chronological ordering (cap consecutive stories per source). No personalized ranking. |
| Publisher visit | Yes — open original |
| Substitutes article? | No — link-out required for full text |
Rights status (four layers)
WeSearch handling by dimension
| Indexing | May the item be indexed (stored, ranked, made findable)? | Allowed |
| Snippet | May a short excerpt of the publisher's text be shown? | Allowed |
| AI summary | May WeSearch generate its own short summary of the article? | Limited |
| Retrieval / RAG | May the content be exposed for third-party retrieval-augmented generation? | Not asserted |
| Model training | May the content be used to train AI models? | Not asserted |
| Commercial reuse | May the content be reused commercially? | Not permitted |
Basis: Derived from the published RSS/Atom feed. Contact: [email protected]. Reviewed: 2026-07-24.
Opening excerpt (first ~120 words) tap to expand
To optimize this track, the performance team implemented a series of algorithmic fusions and precision co-designs:1. Causal Conv1D FusionThe GDN recurrent update is preceded by a causal 1D convolution (K=4). Initially, this was compiled as an independent operation, forcing the intermediate convolution outputs to be written to and read from HBM. We designed a register-level sliding window algorithm that caches historical token states directly within the TPU's VPU registers. Fusing the 1D convolution and the GDN recurrent state update into a single execution block eliminated 6 redundant HBM round-trips (see PR #2823).2. Algebraic Identity OptimizationsWe restructured the linear attention update equations to exploit algebraic identities.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Google Developers Blog.
Discussion
0 commentsMore from Google Developers Blog
