An inference server for AMD GPUs, in C++ and HIP
Comments on “An inference server for AMD GPUs, in C++ and HIP”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Radiance, published on Codeberg by StillDeadcode, is a C++ and HIP inference server for AMD GPUs. It serves language, image, video, speech and music models behind OpenAI-, Anthropic- and llama-server-compatible APIs, with architectures and kernels loaded as plugins. Bundled kernels target RDNA4 cards. The README gives throughput figures, such as 166 tok/s for GLM-5.3-Flash on eight Radeon AI PRO R9700s, but they come from the author's own runs without independent sources, and four-rank speed against two ranks is stated as unmeasured.
Why it ranked
No. 13 of 43 with a score of 0.64. The panel rated it highest on technical depth (0.96) and new information (0.94), lowest on audience breadth (0.44). Placed No. 13, No. 12 by score alone: the fifth AI & machine learning story of the day (−0.040). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Substantial, technically detailed open-source inference server for AMD GPUs with broad API compatibility and many benchmarks; mainly a release with a long, dense README and some promotional links.
- AI judgeJudge 3gpt-5.6-luna
- A highly detailed, reproducible engineering release for AMD GPU inference, covering kernels, quantization, APIs, deployment, benchmarks, plugins, testing, and licensing. It is highly actionable but primarily relevant to AI infrastructure and AMD hardware users.
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
0.64total under rubric v2.3
- Significance0.57
- Reader interest0.72
- Audience breadth0.44
- Practical relevance0.79
- Substantiated0.81
- Technical depth0.96
- New information0.94
- Counts against: Promotional0.15
- Counts against: Overclaim0.17
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? | 0.59 | 0.45 | 0.68 | 0.57 | 0.23 | +0.24 | +0.138 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.67 | 0.60 | 0.88 | 0.72 | 0.28 | +0.26 | +0.186 |
| Audience breadthHow many people who build technology does it concern? | 0.43 | 0.40 | 0.48 | 0.44 | 0.08 | +0.13 | +0.057 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.84 | 0.70 | 0.82 | 0.79 | 0.14 | +0.22 | +0.173 |
| 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. | 0.93 | 0.60 | 0.90 | 0.81 | 0.33 | +0.05 | +0.040 |
| Technical depthHow technically substantive is it? | 0.99 | 0.90 | 0.98 | 0.96 | 0.10 | +0.05 | +0.048 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.92 | 0.90 | 0.99 | 0.94 | 0.09 | +0.05 | +0.047 |
| 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. | 0.05 | 0.35 | 0.04 | 0.15 | 0.31 | −0.19 | −0.028 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.14 | 0.30 | 0.08 | 0.17 | 0.22 | −0.11 | −0.019 |
| Sum of contributions | +0.642 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.642 | ||||||
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/2026-11-02/13/"><img src="https://rubric.hrih.si/badge/2026-11-02/13.svg" alt="Ranked #13 on Rubric, 2 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
On two R9700s, the w4a8 Gemma 4 31B container converted by the author decodes at 103 tok/s on prose and 141 on code, versus 112 and 150 for keptn42's container, and accepts 0.46 of its drafts on prose against 0.50.
The article gives no link or citation for it.
The article says “On two R9700s that container decodes at 103 tok/s on prose and 141 on code, against 112 and 150 for keptn42's, and accepts 0.46 of its drafts on prose against…”
The source says No source to compare with.
No source givenBenchmark result
On two R9700s, TranslateGemma 27B decodes 56 tok/s a stream (205 at four) and scores mean KL 0.042 against the bf16 model on held-out translations; with a DFlash2 drafter, 152 a stream.
The article gives no link or citation for it.
The article says “On two R9700s it decodes 56 tok/s a stream (205 at four) and scores mean KL 0.042 against the bf16 model on held-out translations; with a DFlash2 drafter, 152 a…”
The source says No source to compare with.
No source givenNumber
On eight R9700s, GLM-5.3-Flash prefills 3958 tok/s at 6k tokens, decodes 166 tok/s on one stream and 426 over eight, and 506 under eight streams of agent turns.
The article gives no link or citation for it.
The article says “On eight R9700s it prefills 3958 tok/s at 6k tokens, decodes 166 tok/s on one stream and 426 over eight, and 506 under eight streams of agent turns”
The source says No source to compare with.
No source givenNumber
On two R9700s a 2048x2048 Qwen-Image picture takes 61 s at the base model's 40 steps and 13 s with Turbo; 1024x1024 takes 6 s and 2.6 s.
The article gives no link or citation for it.
The article says “On two R9700s a 2048x2048 picture takes 61 s at the base model's 40 steps and 13 s with Turbo (56 and 12 s with --tp-wire wht6); 1024x1024 takes 6…”
The source says No source to compare with.
No source givenNumber
On two R9700s a 5.17 s clip at 1344x768 takes about 68 s with FastH3 and 8.8 minutes at MiniMax-H3's 50 steps; with the latent upscaler MiniMax-H3's clip takes 3.4 minutes.
The article gives no link or citation for it.
The article says “On two R9700s a 5.17 s clip at 1344x768 takes about 68 s with FastH3 and 8.8 minutes at MiniMax-H3's 50 steps”
The source says No source to compare with.
No source givenNumber
Breeze-TTS-2 makes speech in 0.6 of its playing time on one card, with first streamed audio 0.21 s after the request.
The article gives no link or citation for it.
The article says “made in 0.6 of its playing time on one card, its first streamed audio 0.21 s after the request”
The source says No source to compare with.
No source givenNumber
The Docker image's compiler keeps VOPD pairing, which is 1.4-3.5% faster on the fp8 prefill GEMMs.
The article gives no link or citation for it.
The article says “keeps the pairing, which is 1.4-3.5% faster on the fp8 prefill GEMMs”
The source says No source to compare with.
No source givenGeneralization
The per-step rate of four ranks against --tp 2 on the same cards has not been measured.
The article gives no link or citation for it.
The article says “Four ranks is a capacity configuration: its per-step rate against --tp 2 on the same cards has not been measured”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 2 Nov 2026, 05:46 UTC. It played no part in the ranking.