Busbar

RonanLabsAI/dgx-station-frontier

Receipts-first serving recipes and measurements for two NVIDIA DGX Station GB300s

Visit GitHub

Python

2 contributors

Apache License 2.0

FeaturedAdvertise to 180K+ monthly visitors

Bifrost

Enterprise AI Gateway Built For Scale: Route, Govern, and Secure All Your AI Traffic

ClawBench

a benchmark for personal agents

Busbar

Control what AI can do before it acts

Momentum

How fast RonanLabsAI/dgx-station-frontier is gaining stars, and where it lands on Trendshift and GitHub Trending

Activity and trending

Stars, forks and merged pull requests each day or month, lined up with the repository's rank on Trendshift Trending

Oct 8 – Oct 23, 2026

Stars +15 in 16 days

Forks +0 in 16 days

Merged PRs 0 in 16 days

Trendshift Trendingwhere the repository ranked each day it made the top 25

Not in the Trendshift top 25 in this period

Oct 8Oct 10Oct 12Oct 14Oct 16Oct 18Oct 20Oct 22

Stars, forks, merged pull requests and Trendshift Trending rank, by day
Day	Stars gained	Forks gained
Thu, Oct 8, 2026	+0	+0
Fri, Oct 9, 2026	+0	+0
Sat, Oct 10, 2026	+1	+0
Sun, Oct 11, 2026	+1	+0
Mon, Oct 12, 2026	+12	+0
Tue, Oct 13, 2026	+0	+0
Wed, Oct 14, 2026	+0	+0
Thu, Oct 15, 2026	+0	+0
Fri, Oct 16, 2026	+1	+0
Sat, Oct 17, 2026	+0	+0
Sun, Oct 18, 2026	+0	+0
Mon, Oct 19, 2026	+0	+0
Tue, Oct 20, 2026	+0	+0
Wed, Oct 21, 2026	+0	+0
Thu, Oct 22, 2026	+0	+0
Fri, Oct 23, 2026	+0	+0
Total	+15	+0

#1 on the list
Rank that day
Not trending

Social mentions

Recent discussions about this repository across the web

Oct 141 post
- I wrote predictions down first. One missed: accepted length was 2.74, predicted 2.8 to 3.0. Random-token prompts gave 1.5x only under greedy decoding; with the repo bench's default sampling there is…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
Oct 131 post
- Prefill and decode want different expert maps. On one DGX Station, DeepSeek-V4.1-Flash read 16K prompts 7% faster, decode and long-context accuracy unchanged. Push the map further: +34% prefill, but…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
Oct 113 posts6 posts
- Four days, two DGX Station GB300s, five bugs. Every one looked like a model problem. One was in a GPU kernel; the others hid in the compiler, the inference engine, the OS and an overnight driver…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
- One GB300, one GPU: - 173 tok/s for one user with the DFlash2 drafter, 92 without - 404 tok/s total with 8 users - ~4,500 tok/s reading 8K and 32K prompts - GSM8K 49/50, drafted answers identical to…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
- Three bugs we hit on two DGX Stations, all now filed upstream: 1. FlashInfer b12x 4-bit MoE summed rows it never wrote: NaN 2. Same kernels: small scales made weights vanish, rows lost while cosine…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
- Same server, same drafter, same settings. DeepSeek-V4-Pro on two DGX Stations ran 1.77x faster on real text and 0.58x as fast on random tokens. One number predicts which you get, and most benchmark…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
- The same drafter made DeepSeek-V4-Pro 1.77x faster on real text and 0.58x as fast on random tokens. Why, how to predict it, and how to read speculative-decoding claims. Announce post (269 of 280…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
- MiMo-V2.6-Pro on two DGX Stations, six memory tiers: 503 tok/s at 16 users on J-M's benchmark, 1.81x the same pair without RTX PRO 6000 sidecars. Not one expert left in slow memory. Built on…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
Oct 102 posts
- Thinking about an RTX PRO 6000 for your DGX Station? On real text it made DeepSeek-V4.1-Flash 2.7-3.25x faster at 16-32 users, for about twice the GPU power. For one user, 1.7-1.9x. When it pays and…
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com
- DeepSeek-V4-Pro at its released precision across two DGX Station GB300s: 83-95 tok/s for one user and 274-313 for 16 on public text, step by step, including the hang.
  Read full on X, posted by @SJRonanMD
  @SJRonanMD · x.com