GAAC: Lossless KV Cache Compression for Faster LLM Inference
Comments on “GAAC: Lossless KV Cache Compression for Faster LLM Inference”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Researchers at the University of Notre Dame describe GAAC, a lossless KV cache codec that stores BF16 exponents as short offsets from calibrated per-channel bases and reconstructs them inside a fused attention kernel. It compresses the cache about 1.41× across five models and cuts resident KV memory by roughly 26%. On DGX Spark, 64K-context decode attention runs 1.19–1.27× faster than FlashAttention-2, and serving throughput rises 13–29% at equal KV memory. These are the authors' own measurements, and the Apple M4 speedups are smaller at short contexts.
Why it ranked
No. 12 of 59 with a score of 0.65. The panel rated it highest on technical depth (0.93) and new information (0.91), lowest on audience breadth (0.55). Placed No. 12, No. 15 by score alone: the fourth AI & machine learning story of the day (−0.020). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Substantive lossless KV-cache compression paper with concrete benchmarks and kernel details on LLM inference efficiency; relevant to builders, though a niche technique with some unverified serving-level claims.
- AI judgeJudge 3gpt-5.6-luna
- A detailed original study introduces a lossless KV-cache codec with fused kernels, serving integration, and measurements across NVIDIA and Apple hardware. It is highly actionable for inference engineers, though results are limited to selected models and platforms.
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
0.65total under rubric v2.3
- Significance0.57
- Reader interest0.70
- Audience breadth0.55
- Practical relevance0.70
- Substantiated0.85
- Technical depth0.93
- New information0.91
- Counts against: Promotional0.05
- 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? | 0.58 | 0.40 | 0.74 | 0.57 | 0.34 | +0.24 | +0.137 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.59 | 0.60 | 0.90 | 0.70 | 0.31 | +0.26 | +0.181 |
| Audience breadthHow many people who build technology does it concern? | 0.64 | 0.40 | 0.62 | 0.55 | 0.24 | +0.13 | +0.072 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.73 | 0.60 | 0.78 | 0.70 | 0.18 | +0.22 | +0.155 |
| 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 | 0.75 | 0.84 | 0.85 | 0.21 | +0.05 | +0.043 |
| Technical depthHow technically substantive is it? | 0.97 | 0.85 | 0.96 | 0.93 | 0.12 | +0.05 | +0.046 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.91 | 0.85 | 0.98 | 0.91 | 0.13 | +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. | 0.08 | 0.05 | 0.01 | 0.05 | 0.07 | −0.19 | −0.009 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.22 | 0.25 | 0.12 | 0.20 | 0.13 | −0.11 | −0.022 |
| Sum of contributions | +0.649 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.649 | ||||||
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/2026-11-01/12/"><img src="https://rubric.hrih.si/badge/2026-11-01/12.svg" alt="Ranked #12 on Rubric, 1 Nov 2026" width="238" 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
GAAC achieves 1.405–1.407× lossless BF16 KV compression across five tested models.
The article gives no link or citation for it.
The article says “GAAC reaches a compression ratio of 1.405–1.407 $\times$ on every model”
The source says No source to compare with.
No source givenBenchmark result
On NVIDIA DGX Spark, GAAC accelerates decode attention at a 64K-token context by 1.19–1.27× compared with FlashAttention-2 on BF16 KV.
The article gives no link or citation for it.
The article says “GAAC is 1.19–1.27 $\times$ faster on all four models”
The source says No source to compare with.
No source givenNumber
GAAC speeds up end-to-end W4A16 decode steps by 1.04–1.15× on every model and context length at batch size 1.
The article gives no link or citation for it.
The article says “GAAC speeds up every model at every context length by 1.04–1.15 $\times$”
The source says No source to compare with.
No source givenNumber
Under comparable KV-memory budgets, GAAC raises aggregate serving throughput by 13–29%, with the largest gains on Qwen3.8-27B (+27%) and Qwen3.6-35B-A3B (+29%).
The article gives no link or citation for it.
The article says “Under the same budget, GAAC fits one more sequence for every model, and aggregate throughput rises by 13–29%”
The source says No source to compare with.
No source givenNumber
GAAC reduces resident KV memory by approximately 26%, for example from 19.3 GB to 14.2 GB for Qwen3-8B at 128K.
The article gives no link or citation for it.
The article says “In all cases GAAC also cuts resident KV memory by about 26%, for example from 19.3 GB to 14.2 GB for Qwen3-8B at 128K”
The source says No source to compare with.
No source givenBenchmark result
On Apple M4, GAAC's Metal attention kernel is 1.02–1.25× faster than a BF16-KV baseline across all ten configurations.
The article gives no link or citation for it.
The article says “the GAAC attention kernel is 1.02–1.25 $\times$ faster than the BF16-KV baseline across all ten configurations”
The source says No source to compare with.
Couldn’t checkCausal conclusion
SplitZip's format cannot be decoded piece by piece inside attention because it is designed to decompress the full cache once after transfer.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “it is designed to decompress the full cache once after transfer, and its format cannot be decoded piece by piece inside attention”
The source says No passage could be compared.
Source cited by the article: doi.org/10.48550/arXiv.2605.01708
No source givenNumber
Adding a fourth 2-bit mode improves compression ratio by only 2.3% while increasing decoding latency to 2.8 times that of the 3+4+raw design.
The article gives no link or citation for it.
The article says “adding this mode improves the compression ratio by only 2.3% while increasing decoding latency to 2.8 times that of the 3+4+raw design”
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 1 Nov 2026, 05:47 UTC. It played no part in the ranking.