Dnotitia Unveils STAR-KV Compression Method Selected as ICML 2026 Spotlight Paper

Jul 2, 2026
Dnotitia introduced STAR-KV, a cache compression method that reduces key-value cache size by up to 20 times and speeds up generation throughput by over three times. The research, developed with UC San Diego, has been recognized as a Spotlight paper at ICML 2026 in Seoul.

Dnotitia Inc. has released the paper and source code for STAR-KV, a key value cache compression method that achieves up to 20 times reduction in cache size and faster inference speeds, announced in a press release. The research was conducted in collaboration with UC San Diego's VVIP Lab.

The paper, titled "STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control," was selected as a Spotlight paper at ICML 2026, which will be held in Seoul from July 6 to 11. Out of 23,918 reviewed papers, 536 received this recognition.

According to Dnotitia, STAR-KV reduces key value cache size by up to 75 percent using low-rank compression, and by up to 20 times when combined with a mixed precision quantization method. The technique also uses custom GPU kernels to improve attention computation speed by up to 6.9 times and overall generation throughput by up to 3.1 times.

Dnotitia plans to further develop STAR-KV for AI service environments and adapt it to open source large language model inference frameworks such as vLLM.

We hope you enjoyed this article

Consider subscribing to one of our newsletters like Silicon Brief or Daily AI Brief.

Also, consider following us on social media:

AI Chips & Datacenters AI Brief AI Brief (X)
Free newsletter

Silicon Brief

Weekly coverage of AI hardware developments including chips, GPUs, cloud platforms, and data center technology.

AI Token Webinar

Whitepaper

Tensordyne Napier: What If One Rack Could Do the Work of Nine?

Tensordyne

This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.

Read more

You may also like

Free, six days a week

Daily AI Brief: the AI news that matters, in your inbox.