189 results for GPUs · 2.410s

News for “GPUs”
18 results • 2405 ms server time
Moozonian News
github.com• Aug 12, 2026• 1 min read
Show HN: Tynx - train ONNX models with a PyTorch-shaped API (Tynx is a small, self-contained ONNX runtime with a PyTorch-shaped API for inference and training.The .whl is less than 20MB. GPU execution uses Burn/CubeCL + wgpu enabling it to run across OSes and GPUs, without requiring any extra libraries. $ pip install tynx And API import tynx as tx model = tx.nn.Sequential( tx.nn.Linear(8, 16), tx.nn.ReLU(), tx.nn.Linear(16, 2) ) optimizer = tx.optim.Adam(model.parameters(), lr=1e-3) loss = tx.nn.functional.cross_entropy(model(x), target) loss.backward() optimizer.step() The runtime is written in Rust, and also can compile to the browser(PoC done).It’s early, I'd love for feedback and possible use cases & API expansion where this could be beneficial.
Moozonian News
news.ycombinator.com• Aug 10, 2026• 1 min read
Computers Are Just a ToolComputers are just a toolThey are there when you need them. They get the job done (very) well. They look sleekYou put them away when you want to connect. Be present. Listen. Engage. RelateThey do not ask for attention. They do not desire attention. Would a wrench ask you how your day went? Would your calculator? Why should your “AI problem solver”?Internet 1.0 was a bubble. Nasdaq IPOs liquidating in 6 months. Some cool things were discovered and built in the aftermath, including Alphabet’s theory of managementInternet 2.0 was a stretched out disaster. Engagement farming. Addiction loops built using dark patterns to capture your spend. Venture deployed to sales and marketing not R&D.Popularized by: Facebook Zynga Slack a16z New Twitter (not old twitter)Internet 3.0 whispered decentralization but became a centrally controlled casino. Crypto, kalshi, polymarket. Why run GPUs on CNG to validate your bet’s outcome when you can gamble on zero carbon? Innovation! They didn’t even try to hide
Moozonian News
news.ycombinator.com• Jul 25, 2026• 1 min read
Programmer GovernmentHey I had an idea. What if we created the bitcoin for government?So like a king sucks because it's the equivalent of one GPU.A republic is better because it's like a big GPU cluster of hundreds of GPUs.But what if we made like a bitcoin like distributed governance system that was like, I don't know, a million GPUs?We have the internet which is like infinite bandwidth compared to human needs, so we have the bandwidth to link all our minds together.So what if we did this. We had a simple program that essentially each person could host a node and get 5 random people's data, so like we would store data on numerous nodes for redundancy.We have enough gigabytes, we could store a routing table for everybody's data in America since there's only like 300 million people, if each person had a kilobyte of data, it could be stored in 300 GB which is a lot but actually still doable on consumer level hardware.So then we would have a routing table of which nodes has which person's data, and then we co
Advertisement
Moozonian News
watchmachinego.com• Jul 23, 2026• 1 min read
Show HN: WatchMachineGo – A visualizer to show hardware performing LLM inferenceI love learning new concepts by playing around with them (think explorable explanations (https://en.wikipedia.org/wiki/Explorable_explanation), and I got more and more interested in running local LLMs on my own hardware. Since we live in the age of "speak, and your side project will be done" (joking, but LLMs are really helpfull for me when it comes to frontend stuff and for brainstorming in general!), I built the tool I needed myself :)WatchMachineGo is basically a visual simulator that simulates how a local LLM model gets loaded, prefilled and then used for inference, while showing the effects of different hardware parameters like memory bandwitch or setups like no GPU, two GPUs and so on.All free and without ads, forever.I plan to open source it too, but want to think about how first, still.Note: It is still under construction!If you check it out: Thank you very much and I hope that it will be time well spent! :)
Moozonian News
openparser.dev• Jul 23, 2026• 1 min read
Show HN: Hosted PaddleOCR-VL-1.6 APIWe do a lot of document processing, and PaddleOCR-VL-1.6 is one of the best open models we've used. We wanted to use it as an api, but couldn't find anyone hosting it. so we started hosting it ourselves, but to get full utilization on the GPUs, we decided to let others use it as well.If we did the pricing math right (which I hope we did), it should be the cheapest OCR there is, at just $0.001/page. Let us know if you find it useful
Moozonian News
github.com• Jul 17, 2026• 1 min read
Show HN: Open-source verification and tuning layer for self-hosted LLMsHi HN,Thomas here, I'm a Machine Learning, and I've been deploying models in prod for the last 5 years and decided it was time to open-source some of my learnings!In Coloma you'll find my vLLM tuner, I use it to ensure models don't OOM at runtime, and are optimized for the metric that matters the most for that deployment: low per-user latency vs maximum system throughput. To make it easier the dashboard ships with an estimator of the performance expected at runtime depending on the parameters picked.I've also started to add my outputs verification system that triggers when the GPUs are idling, so that at runtime the predictions can run fast, and get flagged later on for continuous improvement.Really looking forward to see if it that can help someone!
Moozonian News
news.ycombinator.com• Jul 16, 2026• 1 min read
Ask HN: Best way to buy second hand compute (in a future AI crash)So, in a land before kids and mortgages I had a homelab, and want to start again.I also think that give it a few months and the AI bud out bubble will burst and a couple of data r free worth of GPUs might come availableBuying off eBay sounds crazy so what’s the right way to find compute from sell offs, refreshed estates and so on?
Moozonian News
github.com• Jul 16, 2026• 1 min read
Show HN: Trellis2.c – Local 3D generation with Vulkan and CUDAHi HN, I started this because I wanted something in the spirit of llama.cpp and stable-diffusion.cpp, but for local 3D generation model: native executables, no Python/PyTorch runtime, and both CUDA and Vulkan backends. I’ve mainly tested it on an NVIDIA RTX 4090. I’m particularly interested in Vulkan results on AMD and Intel GPUs, where driver behavior and memory allocation may differ. If you try it, GPU, driver, OS details, and failure logs would be extremely helpful. The main limitation right now is performance. The implementation has not yet been heavily optimized, so inference is currently slower than mature, fully optimized PyTorch or Vulkan backends. There is still substantial room for kernel, memory-use, and scheduling improvements.
Moozonian News
usefeyn.com• Jul 6, 2026• 1 min read
Show HN: Pulpie – Models for Cleaning the WebHey HN, I'm Shreyash, founder of Feyn. We built Pulpie, a family of Pareto optimal models for cleaning the web. Pulpie strips boilerplate (ads, footers, sidebars) from raw HTML and returns just the main content as HTML or Markdown.We match SOTA extraction quality while being 20x cheaper. Cleaning 1 billion webpages costs $7,900 with Pulpie versus $159,000 with Dripper, the current leading extractor.The gains come from architecture. Today's leading extractors are decoders that generate output one token at a time. Each step reads the full model from memory to produce a single token. Conversely, Pulpie models are encoders. They run one forward pass over the full input HTML and label each block as boilerplate or content. As a result, Pulpie is compute-bound while decoders are memory-bound. Cheaper GPUs have relatively more compute than memory bandwidth. This makes Pulpie easy to run optimally.Here's Pulpie and Dripper cleaning the same pages side by side: https://www.youtube.com/watch?v=ib
Advertisement
Moozonian News
github.com• Jun 11, 2026• 1 min read
Show HN: Run production AI in your cloud in 5 minsHello! I’m Yodi from Dagploy.com.I spent the past 12 months building AIOps Infra tools to help anyone run self-hosted AI in their own cloud. Run any open source models and tools. Setup just takes about 5 minutes.Previously, I deployed a AI short-video recommendation system serve 60 million users. It took me 8 months to build the AI Infra for setup GPUs, deal with quota, network problem, disk issue and many more. That painful experience inspired me to build Dagploy, so others can get started much faster.Intro video: https://www.youtube.com/watch?v=BCgOexZ_fr8Website: https://www.dagploy.comI’d love to hear your feedback, questions, or suggestions.
Moozonian News
bing.com• Jun 5, 2026• 1 min read
How Anthropic, OpenAI And Nvidia Are Defining The AI EconomyThis voice experience is generated by AI. Learn more. This voice experience is generated by AI. Learn more. Nvidia has dominated of the market for the most advanced GPUs to run enterprise AI functions ...
Moozonian News
news.ycombinator.com• May 17, 2026• 1 min read
Project Prism |Fullstack Engineer – Abu Dhabi (Onsite) – Full-Time – Presight.aiPresight.ai is a publicly listed company with various projects in the field of big data analysis and ML models application. Our solutions work domestically and internationally.For one of such projects we are looking for experts in Frontend development, utilizing Typescript, React, Mobx.Backend features Node.js and Elasticsearch as main storage, plus multiple services powered by ML models accelerated on GPUs (incl. RAG and Agentic based analysis).This product specialises on sifting through large media and text content, capturing trends, finding answers and is loved by corporates. UI rapidly changes and we want to find several people who can help us stay mainstream.Will be glad to welcome you to the multicultural team of like minded people.You can reach me on LinkedIn and DM: https://www.linkedin.com/in/yury-lapin Or email me on yury.lapin@presight.ai with you CV :-)
Moozonian News
news.ycombinator.com• Apr 7, 2026• 1 min read
Elon Musk, Quantum Microtubules, and the Race for the Conscious MachineA recent Quanta Magazine piece by John Pavlus examines a burning question: how close are humanoid robots to becoming more sophisticated than humans? While neural networks on fast GPUs have turbocharged computer vision and reinforcement learning, allowing robots to perceive environments better, a massive gap remains between "moving" and "being."Engineers have moved beyond the "linear inverted pendulum" model, using deep reinforcement learning to act as whole-body controllers. As Pulkit Agarwal puts it: “To have robots which work like humans, I think we have to master physics.” This means moving beyond maps and truly understanding force and inertia.At Google DeepMind, Carolina Parada emphasizes VLA (Vision-Language-Action)—giving a robot one cohesive "brain" instead of three that don't get along. Meanwhile, Jonathan Hurst focuses on Quasi-Direct Drive (QDD) motors—robotic "muscles" that balance strength with sensitivity.Yet, Russ Tedrake argues the bodies are already good enough; the pro
Moozonian News
news.ycombinator.com• Apr 5, 2026• 1 min read
Show HN: 1B EmbeddingsWe built a vector search engine based on Quantized Tensor Train (QTT) decomposition. Instead of approximate nearest neighbor (ANN) indices like HNSW or IVF, we factorize the entire dataset into a compressed tensor format and serve exact cosine similarity queries directly from the compressed representation. The headline: 1 billion vectors on a single H100, 38ms query, 100% recall, 66 GB serving.Recall improves with scale at fp16: 96% at 400M → 98% at 500M → 99% at 600M → 100% at 1B. This is the opposite of ANN indices, where recall degrades with scale. More data helps the decomposition converge.Every number below is measured, not projected. Full benchmark suite across 4 GPUs at 3 precision tiers. H100 80GB, 384-dim embeddings, rank=32.fp16 (Scale tier) — H100 80GB: 100M: 5.87ms p50, 6.6 GB serving, 100% recall, 46.5x compression 500M: 20.54ms p50, 33.0 GB serving, 98% recall, 46.5x compression 1B: 38.51ms p50, 66.0 GB serving, 100% recall, 46.5x compression fp32 (Production tier) — H100
Advertisement
Moozonian News
news.ycombinator.com• Mar 26, 2026• 1 min read
A 100% serverless RAG that extracts complex tables better than NotebookLMStandard RAG wrappers completely mangle structured data.Here is Google's NotebookLM vs. KRAG pipeline extracting the exact same nested tables: https://www.youtube.com/watch?v=sngsQ6HJnlYNotebookLM hallucinates rows. KRAG extracts the grid flawlessly.Under the hood:100% Serverless: Deploys instantly (Bring Your Own Keys, pay your own compute).100% Private: Built using slms which deploy serverlessly on modal gpus.Enterprise Grade: 3 native layers of encryption.100% Open Source & FreeGithub: https://github.com/SaurabhJain708/krag