Training a 1T-Params MoE on EU public infrastructure
Comments on “Training a 1T-Params MoE on EU public infrastructure”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Domyn, Fraunhofer IAIS and Fraunhofer IIS describe training a roughly 1-trillion-parameter Mixture-of-Experts model on Jupiter, a Hopper-based EuroHPC supercomputer built for scientific workloads, using only user-level access.
Why it ranked
No. 65 of 137 with a score of 0.62. The panel rated it highest on technical depth (0.98) and new information (0.92), lowest on audience breadth (0.62). The claim check took 0.100 off. Placed No. 65, No. 53 by score alone: the 19th AI & machine learning story of the day (−0.080). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Detailed, reproducible engineering writeup on training a trillion-parameter MoE on Jupiter with concrete throughput measurements and upstream contributions; highly technical and novel, though audience is narrower than general tech.
- AI judgeJudge 3gpt-5.6-luna
- A highly detailed, measured engineering account of training a trillion-parameter MoE on shared European supercomputers, including hardware constraints, parallelism, memory behavior, kernels, communication, and substantial upstream fixes.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.62total under rubric v2.3
- Significance0.72
- Reader interest0.86
- Audience breadth0.62
- Practical relevance0.71
- Substantiated0.90
- Technical depth0.98
- New information0.92
- Counts against: Promotional0.10
- Counts against: Overclaim0.31
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.72 | 0.55 | 0.90 | 0.72 | 0.35 | +0.24 | +0.173 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.77 | 0.85 | 0.95 | 0.86 | 0.18 | +0.26 | +0.223 |
| Audience breadthHow many people who build technology does it concern? | 0.55 | 0.60 | 0.70 | 0.62 | 0.15 | +0.13 | +0.080 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.68 | 0.60 | 0.85 | 0.71 | 0.25 | +0.22 | +0.156 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.96 | 0.80 | 0.95 | 0.90 | 0.16 | +0.05 | +0.045 |
| Technical depthHow technically substantive is it? | 1.00 | 0.95 | 1.00 | 0.98 | 0.05 | +0.05 | +0.049 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.90 | 0.90 | 0.95 | 0.92 | 0.05 | +0.05 | +0.046 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.07 | 0.15 | 0.08 | 0.10 | 0.08 | −0.19 | −0.019 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.53 | 0.20 | 0.20 | 0.31 | 0.33 | −0.11 | −0.034 |
| Sum of contributions | +0.719 | ||||||
| Claim-check penalty | −0.100 | ||||||
| Total, as published | 0.619 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-13/65/"><img src="https://rubric.hrih.si/badge/2026-11-13/65.svg" alt="Ranked #65 on Rubric, 13 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 1 differs from source · 6 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
Domyn's changes raised training throughput from 90 to ~1,300 tokens/s per GPU on 384 GH200 GPUs (96 nodes).
The article gives no link or citation for it.
The article says “the changes that raised throughput from 90 to ~1,300 tokens/s per GPU on 384 GH200 GPUs (96 nodes”
The source says No source to compare with.
Couldn’t checkFirst, only or fastest
Jupiter is the first European supercomputer in the exascale class, with a Booster module of 5,884 compute nodes each with four NVIDIA GH200 superchips.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Jupiter is the first European supercomputer in the exascale class. Its Booster module has 5,884 compute nodes, each with four NVIDIA GH200 Grace Hopper Superchips”
The source says No passage could be compared.
Source cited by the article: apps.fz-juelich.de/jsc/hps/jupiter/configuration.html
No source givenNumber
The consortium was given access to 2.5% of the EuroHPC compute capacity on public European supercomputers.
The article gives no link or citation for it.
The article says “The consortium was given access to 2.5% of the EuroHPC compute capacity on public”
The source says No source to compare with.
Differs from sourceNumber
Moving Newton-Schulz to the GPU reduced the optimizer step from 41.7 s to 2.2 s and raised throughput from 441 to 597 tokens/s per GPU (+35%).
The cited source says something else; both quotes are shown.
The article says “the optimizer step took 2.2 s, which led throughput to rise from 441 to 597 tokens/s per GPU (+35%”
The source says “In my test, with cpu offloading, allowed the optimizer step to get to ~1s from ~15s per step”
Source cited by the article: github.com/pytorch/torchtitan/pull/5352
No source givenNumber
Offline attention-cost balancing cut the most expensive rank's attention work from a median of 7.9× the mean to 1.15×, and reduced step time from 103.8 s to 91.7 s (13% throughput gain).
The article gives no link or citation for it.
The article says “the most expensive rank in a step did a median of 7.9× the mean attention work before balancing, and 1.15× after”
The source says No source to compare with.
No source givenNumber
Gathering expert weights in FP8 cut all-gather time from 53.6 to 35.1 s per step and raised throughput from 763 to 898 tokens/s per GPU (+18%).
The article gives no link or citation for it.
The article says “all-gather time on the GPU decreased from 53.6 to 35.1 s per step; throughput rose from 763 to 898 tokens/s per GPU (+18%”
The source says No source to compare with.
No source givenNumber
Capping NCCL_MAX_CTAS at 16 raised throughput from 1,210 to 1,293 tokens/s per GPU (+7%).
The article gives no link or citation for it.
The article says “With this environment variable alone, throughput rose from 1,210 to 1,293 tokens/s per GPU (+7%”
The source says No source to compare with.
No source givenGeneralization
The final configuration reaches ~1,300 tokens/s per GPU, 14× the baseline, demonstrating that a shared Hopper-generation public system with user-level access can host 1T-scale training.
The article gives no link or citation for it.
The article says “we demonstrated that a publicly funded, shared, Hopper-generation system with user-level access can host training at the 1T scale”
The source says No source to compare with.
Also covered by
Elsewhere
Elsewhere: HN 4 pts, not on front page
Recorded for context only, checked 15 Nov 2026, 20:42 UTC. It played no part in the ranking.