GPU says 35% busy. Busy doing what?
Comments on “GPU says 35% busy. Busy doing what?”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Tensward, which builds an open-source LLM inference tuning tool, found that nvidia-smi's 35% "busy" figure on an H100 running llama.cpp hid a 643 MB per-request copy of conversation state off the GPU for a hybrid Qwen model. Disabling prompt caching and checkpoints raised throughput from 58 to 110 tokens per second, but it slowed a returning-conversation workload on an 8B model by about 20%. The figures are the company's own and unsourced; its CUPTI-based counter is research tooling, and its crash rate was only measured on one H100.
Why it ranked
No. 10 of 60 with a score of 0.68. The panel rated it highest on new information (0.90) and technical depth (0.90), lowest on significance (0.52).
- AI judgeJudge 2claude-haiku-5-5
- Detailed, measured engineering write-up on LLM inference GPU utilisation with reproducible flags and numbers; practical for builders, narrow audience, modest significance, some self-promotion of their tool.
- AI judgeJudge 3gpt-5.6-luna
- A well-supported engineering investigation measures hidden GPU memory-copy costs, explains CUPTI-based instrumentation, reports reproducible performance effects and limitations, and offers actionable diagnosis for LLM inference operators.
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
0.68total under rubric v2.3
- Significance0.52
- Reader interest0.82
- Audience breadth0.53
- Practical relevance0.81
- Substantiated0.88
- Technical depth0.90
- New information0.90
- Counts against: Promotional0.11
- Counts against: Overclaim0.20
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? | 0.52 | 0.35 | 0.68 | 0.52 | 0.33 | +0.24 | +0.124 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.74 | 0.82 | 0.90 | 0.82 | 0.16 | +0.26 | +0.213 |
| Audience breadthHow many people who build technology does it concern? | 0.45 | 0.50 | 0.65 | 0.53 | 0.20 | +0.13 | +0.069 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.81 | 0.80 | 0.82 | 0.81 | 0.02 | +0.22 | +0.178 |
| 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 | 0.88 | 0.88 | 0.16 | +0.05 | +0.044 |
| Technical depthHow technically substantive is it? | 0.99 | 0.88 | 0.83 | 0.90 | 0.17 | +0.05 | +0.045 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.85 | 0.90 | 0.96 | 0.90 | 0.11 | +0.05 | +0.045 |
| 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. | 0.04 | 0.20 | 0.08 | 0.11 | 0.16 | −0.19 | −0.020 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.18 | 0.30 | 0.12 | 0.20 | 0.18 | −0.11 | −0.022 |
| Sum of contributions | +0.677 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.677 | ||||||
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/2026-11-16/10/"><img src="https://rubric.hrih.si/badge/2026-11-16/10.svg" alt="Ranked #10 on Rubric, 16 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 7 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
nvidia-smi reported the H100 as 35% busy while the LLM server delivered 58 tokens per second and users waited almost two seconds for a first token.
The article gives no link or citation for it.
The article says “It said 35% busy. Throughput was 58 tokens per second, and users waited almost two seconds for a first token”
The source says No source to compare with.
No source givenGeneralization
NVIDIA documents GPU utilisation as the share of each sampling period in which at least one kernel was running.
The article gives no link or citation for it.
The article says “NVIDIA documents it as the share of each sampling period in which at least one kernel was running”
The source says No source to compare with.
No source givenNumber
With llama.cpp and Qwen3.8-27B, each request copied 643 MB from the GPU to host memory, versus 49 MB for a regular 8B model on the same prompts.
The article gives no link or citation for it.
The article says “Every request copied 643 MB from the GPU back to host memory. A regular 8B model on the same prompts copied 49 MB”
The source says No source to compare with.
No source givenCausal conclusion
Disabling prompt cache and context checkpoints raised throughput from 58 to 110 tokens per second and busy from 35% to 88%.
The article gives no link or citation for it.
The article says “Throughput went from 58 to 110 tokens per second, and busy went from 35% to 88%”
The source says No source to compare with.
No source givenVersion
On a newer llama.cpp build (b12246, from 10 November), the same workload ran 67 to 75% faster with default settings than on build b11669 (15 October).
The article gives no link or citation for it.
The article says “With default settings it ran the same workload 67 to 75% faster”
The source says No source to compare with.
Couldn’t checkCausal conclusion
Three merged pull requests between the two llama.cpp builds changed how state is copied and stored.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Between the two builds, three merged pull requests changed how state is copied and stored: #30881, #31391 and #30137”
The source says No passage could be compared.
Source cited by the article: github.com/ggml-org/llama.cpp/pull/30881
No source givenNumber
With vLLM on an H100, nvidia-smi read 95 to 97% busy while DCGM SM occupancy was 8 to 12%.
The article gives no link or citation for it.
The article says “nvidia-smi read 95 to 97% busy, while DCGM’s SM occupancy (DCGM_FI_PROF_SM_OCCUPANCY) was 8 to 12%”
The source says No source to compare with.
No source givenNumber
nsys reported 639 MB copied per request, consistent with the census figure.
The article gives no link or citation for it.
The article says “we used it to check our figure (it said 639 MB”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 5 pts, not on front page
Recorded for context only, checked 16 Nov 2026, 05:47 UTC. It played no part in the ranking.