Changelog
Release notes for Dell Deskside Agentic AI Solutions for NVIDIA. Every
performance figure in a release carries provenance through the machine-readable
record in tests/data/latest/sizing.yaml and the raw ladder data under
tests/data/.
1.0.0 — October 20, 2026
First public release: five Solution Templates on five supported Dell deskside platforms powered by NVIDIA, with measured capacity envelopes, executable serving recipes, and operator guides.
Solution Templates
| Template | Category | What it delivers |
|---|---|---|
| Agentic Coding | Horizontal | Private coding assistant — CLI/IDE agents behind an OpenAI-compatible API, with measured per-platform concurrency envelopes |
| Agentic Video Search & Summarization | Vertical | Natural-language search, Q&A, and summarization over local video — LLM + vision-language serving co-located on one system |
| OpenClaw on NemoClaw | Horizontal | The OpenClaw assistant on NVIDIA's NemoClaw runtime — policy engine, OpenShell sandbox, tool-using agents on a locally served LLM |
| Hermes on NemoClaw | Horizontal | Nous Research' Hermes Agent harness on the NemoClaw runtime — served and sandboxed fully on Dell deskside hardware |
| Higher Education AI Lab | Vertical | University AI teaching labs — shared serving tiers, per-course provisioning, seat resets, and a curriculum built on the other templates |
Supported platforms
| Platform | Memory | Tier |
|---|---|---|
| Dell Pro Precision with GB10 (DGX Spark class, aarch64) | 128 GB unified | S · M (1:1 seats) |
| Dell Pro Precision 9 T2 (1× RTX PRO 6000 Blackwell) | 96 GB | S |
| Dell Pro Precision 9 T4 (2× RTX PRO 6000 Blackwell) | 192 GB | M |
| Dell Pro Precision 9 T6 (5× RTX PRO 6000 Blackwell) | 480 GB | L |
| Dell Pro Precision with GB300 (DGX Station class, aarch64) | 252 GB HBM3e | L |
In this release
- Serving recipes (
recipes/): eight pinned, GPU-tuned vLLM recipes across GB10 / Precision / GB300, all sharing one readiness contract — block until healthy, emit a machine-readableREADYline, meaningful exit codes — plus paired health wrappers, a replica launcher, and a LiteLLM router for multi-instance serving. - Measured envelopes: agent- and chat-shape throughput ladders (AIPerf) per platform, per-session concurrency envelopes at explicit floors, and per-workload TP-vs-DP topology data.
- Fleet and offline tooling: Canonical Landscape deployment payloads, air-gap image packaging, LAN multicast distribution for large lab assets, and scripted NemoClaw installation.
- Tokenomics monitoring: Prometheus + Grafana started with every recipe, alert rules, per-session throughput views.
- Architecture views: every Solution Template's guide opens with an SVG architecture diagram (the Higher Education template adds a campus usage and connectivity view), and the interactive Architecture Builder exports any of them as SVG, PNG, or HTML.
- This docs site, generated from the Markdown sources in the repo, with the same navigation on every page and a link to the source repository in the header.
Tagged release: v1.0.0.
Release policy
Version bumps here track the Solution Templates and their measured data: major = breaking changes to the recipe contract or template structure; minor = new templates, platforms, or measured tiers; patch = recipe fixes and corrections. Each entry names what was measured and where the raw data lives.