Kog is going deeper to squeeze more inference out of GPUs | TechCrunch · AuraTracer

SummaryAI generated

On August 14, 2026, TechCrunch reported that Kog is working to further exploit the reasoning potential of GPUs through technical measures. This initiative aims to maximize the utilization of computing resources at the hardware level, thereby improving overall reasoning performance and efficiency.

Related eventsRELATED EVENTS

2026-09-08 10:00“The starting price is over 10,000! The Xiaomi 18 Fold has been released. Lei Jun talks about chip manufacturing: As soon as we start to catch up, Xiaomi will be on the road to victory.”
2026-09-07 13:02“China launched the International Earth-Moon Cube Constellation Program: 30 satellites will be deployed by 2030 to build a global continuous observation network in Earth-Moon space”
2026-09-07 20:42Xiaomi 18 Fold is the first to be equipped with LPDDR6, and Changxin has officially announced that LPDDR6 will be mass-produced for flagship smartphones for the first time

Key entitiesKEY ENTITIES

AMDCerebrasGaël DelalleauKogNvidia

Coverage · reports per dayLANGUAGE SPLIT

Entity relations

SignalsSIGNALS

Keyword heat

Kog1
Cerebras1
AMD1
Nvidia1
Gaël Delalleau1

All reports (1)SOURCES

↗

T TechCrunch 归档 8月 p10 en 2026-08-14 08:00

Kog is going deeper to squeeze more inference out of GPUs | TechCrunch

法国初创公司 Kog 致力于通过软件优化挖掘常规 GPU 的更多推理潜力，其 CEO Gaël Delalleau 表示该软件工程有望成为首个用例。Kog 在五月于 Hacker News 发布技术预览，证明标准数据中心 GPU（如 AMD MI300X 和 Nvidia H200）可实现极快单次请求解码，虽未覆盖笔记本电脑 GPU，但已吸引超过 200 个实质性商业线索。公司计划针对因延迟而放弃 AI 工作流的客户以及依赖提示词生成游戏和应用的设计合作伙伴，目前正专注于加速大模型开发以应对需求，目标是将大型语言模型推理速度提升 30 倍。尽管演示中基于开源 Laneformer 2B 小模型实现了每秒 3,000 个 token 的吞吐量，Delalleau 坚信该软件优化方法同样适用于参数规模更大的 LL…