Six GPUs on One X16 Slot: R9700s and RTX 3090s Behind a PCIe Switch
Comments on “Six GPUs on One X16 Slot: R9700s and RTX 3090s Behind a PCIe Switch”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Doug Calobrisi describes running six GPUs from one x16 slot on a consumer X570 board: four Radeon AI PRO R9700s and two RTX 3090s behind a Broadcom PEX88096 switch, each card on its own x8 link. On the R9700s, Qwen3.8-27B at TP4 reports about 4,700 tokens/s prefill at 32k and 601 tokens/s across sixteen clients. The numbers are the author's own, unsourced measurements on one setup. The build also needed custom kernel modules for BAR sizing.
Why it ranked
No. 21 of 46 with a score of 0.57. The panel rated it highest on technical depth (0.94) and new information (0.92), lowest on audience breadth (0.37). The judges disagreed most on significance (spread 0.42), so the story is marked contested. Placed No. 21, No. 41 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Detailed, specific homelab build with reproducible PCIe, BAR, and driver configuration and benchmarks, but a niche audience and a single-person setup limit significance and breadth.
- AI judgeJudge 3gpt-5.6-luna
- A detailed, measured account of fitting six GPUs behind one PCIe slot, including topology, BAR sizing, kernel workarounds, P2P behavior, and inference benchmarks. Highly actionable for local AI infrastructure builders, though specialized.
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
0.57total under rubric v2.3Contestedjudges differ by 0.42 on significance
- Significance0.44
- Reader interest0.67
- Audience breadth0.37
- Practical relevance0.60
- Substantiated0.79
- Technical depth0.94
- New information0.92
- Counts against: Promotional0.04
- Counts against: Overclaim0.15
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? | 0.49 | 0.20 | 0.62 | 0.44 | 0.42 , above the contested threshold | +0.24 | +0.105 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.64 | 0.55 | 0.82 | 0.67 | 0.27 | +0.26 | +0.174 |
| Audience breadthHow many people who build technology does it concern? | 0.42 | 0.20 | 0.48 | 0.37 | 0.28 | +0.13 | +0.048 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.58 | 0.45 | 0.76 | 0.60 | 0.31 | +0.22 | +0.132 |
| 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. | 0.90 | 0.60 | 0.86 | 0.79 | 0.30 | +0.05 | +0.039 |
| Technical depthHow technically substantive is it? | 1.00 | 0.90 | 0.91 | 0.94 | 0.10 | +0.05 | +0.047 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.88 | 0.90 | 0.98 | 0.92 | 0.10 | +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. | 0.04 | 0.05 | 0.03 | 0.04 | 0.02 | −0.19 | −0.008 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.21 | 0.15 | 0.08 | 0.15 | 0.13 | −0.11 | −0.016 |
| Sum of contributions | +0.567 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.567 | ||||||
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/2026-11-09/21/"><img src="https://rubric.hrih.si/badge/2026-11-09/21.svg" alt="Ranked #21 on Rubric, 9 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 1 matches 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 givenBenchmark result
On four R9700s with TP4, Qwen3.8-27B prefills a 32k-token prompt at about 4,700 tokens per second and a 233k-token prompt at 3,340 tokens per second.
The article gives no link or citation for it.
The article says “Qwen3.8-27B prefills a 32k-token prompt at about 4,700 tokens per second and a 233k-token one at 3,340”
The source says No source to compare with.
No source givenBenchmark result
With sixteen concurrent clients, the four R9700s running radiance deliver 601.1 tokens per second of output in total, versus 151.1 for one client.
The article gives no link or citation for it.
The article says “One client gets 151.1 tokens per second, and sixteen get 601.1 between them”
The source says No source to compare with.
No source givenNumber
All four R9700s copying at once through the switch moved 54.7 GB/s, about twice the roughly 28 GB/s the x16 uplink carries one way.
The article gives no link or citation for it.
The article says “All four copying at once moved 54.7 GB/s, about twice the roughly 28 GB/s the x16 uplink carries one way”
The source says No source to compare with.
Couldn’t checkGeneralization
Bifurcation splits an x16 slot into at most four x4 links, while the switch gives every card its own x8 link.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Bifurcation splits an x16 slot into at most four x4 links, while the switch gives every card its own x8”
The source says No passage could be compared.
Source cited by the article: doug.sh/posts/pcie-bifurcation-switches-retimers
No source givenBenchmark result
On the 3090s with gpu-unlock, a 10.8k-token prompt prefills at about 2,390 tokens per second, and decode runs at 67 tokens per second for one stream and 400 for eight.
The article gives no link or citation for it.
The article says “a 10.8k-token prompt prefills at about 2,390 tokens per second, and decode runs at 67 tokens per second for one stream and 400 for eight”
The source says No source to compare with.
No source givenCausal conclusion
gpu-unlock makes vLLM prefill 13 to 20% faster with no clear decode gain.
The article gives no link or citation for it.
The article says “gpu-unlock makes vLLM prefill 13 to 20% faster with no clear decode gain”
The source says No source to compare with.
Matches sourceCausal conclusion
gpu-unlock removes the GeForce half-rate cap on tensor math with FP32 accumulation.
The cited source says the same thing.
The article says “It removes the GeForce half-rate cap on tensor math with FP32 accumulation”
The source says “gpu_unlock` removes that cap at boot, before the stock NVIDIA driver loads”
Source cited by the article: github.com/Zagitta/gpu-unlock
No source givenCausal conclusion
Turning P2P off roughly halves the 16-client total throughput on the R9700s.
The article gives no link or citation for it.
The article says “Turning P2P off roughly halves the 16-client total”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 9 Nov 2026, 05:46 UTC. It played no part in the ranking.