frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Show HN: Running PrismML's Bonsai inside DRAM by breaking DDR4 timing rules

16•pcdeni•4d ago
The excitement surrounding PrismML’s 1-bit/ternary Bonsai models has the industry closely watching how smartphone giants, particularly Apple, will implement LLMs on edge devices.

Moving AI on-device is a brilliant and necessary strategy. It ensures absolute user privacy in alignment with EU regulations, fundamentally shifts the economics away from costly cloud inference, and paves the way for a significant hardware upgrade supercycle as users seek true AI-capable silicon.

To create a smart on-device "Semantic Router," models need to reach the 27B+ parameter scale. Achieving this on a phone requires extreme quantization, such as PrismML’s ternary weights.

However, a critical hardware reality often overlooked by the software world is that fitting the weights in RAM is not equivalent to moving them. Running a 27B ternary model on standard LPDDR encounters a significant memory bandwidth limitation. Transferring gigabytes of data across the SoC bus for each token generation can lead to thermal throttling of the NPU and excessive battery drain.

This raises an important question: why are we still transferring data to the compute? Why not execute AI inference natively within the memory?

Frustrated with academic PIM simulations that overlook bare-metal physics, I developed CaSA, an architecture that performs ternary LLM inference directly inside COTS DRAM through charge-sharing, completely bypassing the memory bus.

Software quantization is a great initial step, and CaSA provides the physical hardware substrate needed to complete the bridge: https://github.com/pcdeni/CaSA

Comments

SwellJoe•31m ago
Jebus, that is some sloppy prose. Can people not even be bothered to write the summary themselves, anymore? AI doesn't want anything, so they can never have a point of view, so their prose rambles incoherently across all the various prompts they've seen in a project. This project sounds like the ramblings of a crazy person. Even though the fact that DRAM can do any computation is interesting, nobody should have to read this mess.

"why are we still transferring data to the compute? Why not execute AI inference natively within the memory?"

You already answered that question: 47.5 seconds per token from a tiny 2B 1-bit model model.

deivid•27m ago
Interesting project, but the slop readme made me quit reading halfway
butvacuum•25m ago
Very interesting. I don't see it mentioned so I'll ask:

Would being able to alter voltage levels on the fly (eg, cells x y and z get 1.25 while abc get 1.2) expand the ability here?

ilaksh•9m ago
You are saying this is 13 times faster? More proof please. How do we set it up? I really want this to be a real thing.

MAI-Cyber 1

https://microsoft.ai/news/introducing-mai-cyber-1-flash-inside-mdash/
87•migmartri•1h ago•26 comments

Kimi-K3 on HuggingFace

https://huggingface.co/moonshotai/Kimi-K3
1067•nateb2022•12h ago•433 comments

Glue bonds to nonstick surfaces and wipes clean with ethanol

https://cen.acs.org/materials/adhesives/glue-bonds-nonstick-surfaces-wipes-clean/104/web/2026/07
75•gmays•4d ago•33 comments

The computer that helped win World War II

https://spectrum.ieee.org/colossus-computer-ieee-milestone
81•baruchel•4d ago•40 comments

Decathlon Germany adds Wero payment option to decathlon.de website

https://www.sgieurope.com/e-commerce/decathlon-germany-launches-wero-payment-on-its-website/12239...
92•doener•1h ago•77 comments

Show HN: FeyNoBg – Automatic background removal model and training library

https://usefeyn.com/blog/feynobg/
24•snyy•1h ago•6 comments

Exploiting Volvo/Eicher's fleet platform to gain control over all users/vehicles

https://eaton-works.com/2026/07/27/my-eicher-hack/
32•EatonZ•3h ago•5 comments

Removing React.js from the codebase and adapting Htmx for UI interactivity (2023)

https://misago-project.org/t/removing-reactjs-from-the-codebase-and-adapting-htmx-for-ui-interact...
164•Ralfp•8h ago•111 comments

UpCodes (YC S17) is hiring remote AE's to help make buildings cheaper

https://up.codes/careers?utm_source=HN
1•Old_Thrashbarg•1h ago

Show HN: Running PrismML's Bonsai inside DRAM by breaking DDR4 timing rules

16•pcdeni•4d ago•4 comments

Libsm64: Mario 64 as a library for use in external game engines

https://github.com/libsm64/libsm64
126•klaussilveira•8h ago•15 comments

Beckett the Prophet

https://theamericanscholar.org/beckett-the-prophet/
16•samclemens•5d ago•0 comments

Canceling "Hey"

https://chadnauseam.com/random/cancelling-my-hey
120•ChadNauseam•43m ago•17 comments

Magnolias are so old that they're pollinated by beetles, not bees (2025)

https://mymodernmet.com/magnolia-ancient-flowers-beetles/
242•speckx•5d ago•85 comments

Paged Out #9 [pdf]

https://pagedout.institute/download/PagedOut_009.pdf
55•laurensr•3h ago•8 comments

First Robotic Satellite Servicer Launched

https://www.nrl.navy.mil/Media/News/Article/4551871/robotic-servicing-of-geosynchronous-satellite...
51•GlenTheMachine•3d ago•24 comments

Kimi-K3 Technical Report [pdf]

https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf
309•vinhnx•2h ago•115 comments

Should you wash your solar panels?

https://incoherency.co.uk/blog/stories/should-you-wash-your-solar-panels.html
165•surprisetalk•5h ago•137 comments

VLC for Unity now supported on Linux

https://code.videolan.org/videolan/vlc-unity
111•martz•9h ago•34 comments

Building a Fast Lock-Free Queue in Modern C++ from Scratch

https://blog.jaysmito.dev/blog/04-fast-lockfree-queues/
75•ibobev•5d ago•34 comments

How is the Bun Rewrite in Rust going?

https://lockwood.dev/ai/2026/07/27/how-is-the-bun-rewrite-in-rust-going.html
346•tomlockwood•7h ago•267 comments

Towards a Theory of Bugs: The Ruliology of the Unexpected

https://writings.stephenwolfram.com/2026/07/towards-a-theory-of-bugs-the-ruliology-of-the-unexpec...
47•nsoonhui•3d ago•19 comments

Modern email can be built from borrowed parts

https://en.andros.dev/blog/d7ed8b07/modern-email-can-be-built-from-borrowed-parts/
121•andros•9h ago•57 comments

Show HN: Infrawrench – a tool to manage cloud and svcs with workflows and chat

https://infrawrench.com
7•astrid__•1h ago•0 comments

Shay Locomotives

https://www.shaylocomotives.com/
55•Rygian•8h ago•15 comments

The Birth of the American 12-string Guitar (2013)

https://www.harpguitars.net/history/grunewald/12-string.htm
62•bilegeek•7h ago•35 comments

AI companies spend record sums on Washington lobbying

https://www.ft.com/content/d8a5f95e-3b6d-463a-a848-c9ef8e2394db
200•1vuio0pswjnm7•4h ago•97 comments

Elevated errors on Claude Opus 5

https://status.claude.com/incidents/mfdtrknpxghq
79•croemer•6h ago•57 comments

PGSimCity - How PostgreSQL Works

https://nikolays.github.io/PGSimCity/
845•jonbaer•18h ago•82 comments

The Proof Machine (2016)

https://incredible.pm/
25•BenoitP•5h ago•4 comments