frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Show HN: Cuts Long Horizon Inference Costs by 50% via external KV Cache Offload

https://github.com/openlake-project/openlake
21•arnav__1•3h ago
Hey HN, we’re the developers of OpenLake, an open source storage engine for offloading LLM KV caches from GPU memory into a shared tier of RAM and NVMe. We built OpenLake because KV caches are outgrowing GPU memory.

A single 256K token conversation on Gemma 4 31B produces approximately 43GB of KV state, more than half the memory of an 80GB H100. The problem becomes even harder across a cluster: a prefix cached on one GPU host is unavailable when the next request lands on a different GPU, forcing the new GPU to repeat work the fleet has already completed.

Once the KV cache is offloaded, network bandwidth becomes a major constraint on read latency. To move less data across the wire, we built deferred materialization: a custom CUDA kernel that losslessly compresses KV blocks before they leave GPU memory and decompresses them on the GPU after retrieval. In our tests, this achieved:

- 1.72× lossless KV compression. - Approximately 600GB/s decompression throughput on an H100 - 80GB/s of effective KV throughput over a physical 50GB/s link

At 128K context, retrieving cached KV reduces TTFT from 44 seconds to 0.6 seconds, a 66× improvement. Across the complete workload, GPU time reduces from 1,169 seconds to 606 seconds, saving 48.2% of GPU cost.

OpenLake is written in Rust and uses io_uring with one pinned runtime per physical core. We provide connectors for vLLM and SGLang so the cache can be enabled without modifying the inference engine itself.

I would love to hear how others are handling KV reuse across GPU hosts, especially for long contexts, and get to know your thoughts.

Thanks!

GitHub: https://github.com/openlake-project/openlake

Here is our blog: https://cloud.theopenlake.com/blog/taming-the-beast-managing...

Show HN: CheapSecurity – Lightweight, Self-Hosted CCTV for Linux SBCs

https://github.com/gmrandazzo/CheapSecurity
3•zeldone•19m ago•0 comments

Show HN: Cookbook AI – turn scattered recipe files into a print-ready cookbook

https://cookbookai.io/index.html
2•bookletaitcv1re•14m ago•0 comments

Show HN: A client side PDF and article reader backed by Google Drive

https://pdf-reader-peach.vercel.app/
2•northfield27•19m ago•0 comments

Show HN: Reverse Minesweeper

https://sunflowersgame.com/
22•pompomsheep•3h ago•12 comments

Show HN: I mapped every US golf course

https://golfcoursebrowser.com/
184•rickmf•13h ago•112 comments

Show HN: Managing on-premise servers without Kubernetes

https://github.com/ricardoborges/Nautilus
10•r2ob•2h ago•0 comments

Show HN: I built an app for learning and playing Bridge

https://bridge.brannon.online
3•braxxox•1h ago•0 comments

Show HN: A browser-based video editor that renders videos directly with FFmpeg

https://github.com/thiagoalbrecht/weave
3•thiagoas•2h ago•1 comments

Show HN: Brolly, a plain-text weather forecast site

https://brolly.sh/forecast/RWFP2qW8
218•jsax•22h ago•78 comments

Show HN: I made some transistor animations

https://brandonli.net/semisim/animations
203•stunningllama•1d ago•26 comments

Show HN: QueryTuner – SQL diagnostics across 5 DBs, no database connection

https://www.querytuner.com/
4•autoshiftops•4h ago•0 comments

Show HN: SQLiteWatch – profile SQLite on Linux without code changes

https://github.com/Alurith/sqlitewatch
2•imalessandro•2h ago•0 comments

Show HN: CDXE – Cyber Defence Exercise Environment

https://kryptogram.se/cdxe/
3•kryptogram•2h ago•0 comments

Show HN: CrispVoice – Studio voice enhancement that never uploads your voice

https://github.com/Francium-Tech/CrispVoice
5•bragboy•2h ago•0 comments

Show HN: Temporal Context Map

https://ab.neudice.eu/
3•zebike•3h ago•0 comments

Show HN: Wmux – A workspace multiplexer for AI agents

https://github.com/openwong2kim/wmux
6•wong2kim•5h ago•0 comments

Show HN: Virtual time engine to record HTML to 60fps MP4, no drops

https://zlvox.com/tools/html-to-video
2•mrdisloyal•3h ago•0 comments

Show HN: Crossle – A daily puzzle game where you find 3 intersecting words

https://crossle.resistancelabs.tech
3•samurai4869•3h ago•0 comments

Show HN: I wrote this slowly, a needlessly patient text editor

https://i-wrote-this-slowly.fly.dev/
4•pkoird•3h ago•0 comments

Show HN: Lemmings in HTML in Canvas

https://github.com/VorticonCmdr/lemmings
2•VorticonCmdr•3h ago•2 comments

Show HN: Writemark, a dependency free web component for inline Markdown editing

49•_boffin_•19h ago•19 comments

Show HN: Organic cell simulation with infinite zoom (100k cells, raw WebGL)

https://philomatics.com/cell-game-source/
3•cloogshicer•4h ago•1 comments

Show HN: Registrly – cross-referenced UK company and Global LEI lookup

https://registrly.com
3•solodynamo•4h ago•1 comments

Show HN: SpinWin – A macOS menu bar app to visually rotate or spin any window

https://github.com/alokdhir/spinwin
36•dbm5•22h ago•13 comments

Show HN: Peek – Open social media links without distraction

https://github.com/RtiM0/peek
2•otato_potato•5h ago•0 comments

Show HN: Reproducibility Benchmark a Risk Quantitative Model

https://github.com/Fluxara-GOD/fluxara1-audit-verification-pack
11•mingshi_tz•7h ago•0 comments

Show HN: Echo – Fable-level results at 1/3 the cost using open-weight models

472•adam_rida•2d ago•222 comments

Show HN: Manifest – a cheap fingerprint endpoint for detecting page changes

https://omfang.io/
2•Maxnordstrom•5h ago•0 comments

Show HN: Inflect TTS v2+ONNX, 9M/4M text-to-speech models running in the browser

https://inflect-tts.geronimo-labs.com
5•g58892881•9h ago•2 comments

Show HN: A platformer video game using visual representations of functions

https://function-platformer.vercel.app/
2•dbagr•6h ago•0 comments