Alif Semiconductor’s Post

DLSS 3.5 Improved raytracing with "ray reconstruction" Not a ton of info on the tech details of how it works. But with a better ground truth on their AI model in training. They can now achieve a more accurate scene vs the prior implementation which relied more on temporal ray accumulation. Which has great stable results, but can be blurry at times. This new method appears to retain quality much better and with no performance hit as well. Very impressive 👏

Presenting the #Mantra BIONIC Xtreme series - the ultimate MULTI MODEL BIOMETRIC SYSTEM for #Identity Verification & Access Control. Mantra takes pride in presenting our latest embedded #Linux-based device for the cutting-edge Identity Access Management System. Discover the new height of AI-driven innovation with BIONIC Xtreme! BioNIC Xtreme is a powerful edge device that has an AI processor, with GPU & NPU, optimized for deep learning, delivering highly accurate and fast Biometric #Identification. Embrace the emerging revolution! Learn more about BioNIC Xtreme by watching the whole video! #biometrics #biometricsecurity #facialrecognition #identitysecurity #fingerprint #contactless #accesscontrol #identitymanagement #accessmanagement #identityverification #identityandaccessmanagement

🤯 I did not expect #ai could solve #linearalgebra on a locally running #opensource #llm on small GPU less hardware. 🤩 Quality of reasoning & explanations are just amazing https://lnkd.in/ghwX36xf #mathematics #aiadoption #learnbydoing

At NVIDIA #GTC23, check out the latest developments in the newest accelerated framework on the block. It is double exciting for me as my current org - Vector Institute and previous org Shell are both represented in this talk !! Vector Institute's very own Matthew Choi will be talking about our work on benchmarking and scaling #llms using #JAX Register now to see what's in store for JAX support on GPUs, and learn how to provide the latest performance and capabilities for #GenerativeAI. #AI #ML #acceleration https://lnkd.in/dY_YA3MH

Save the date: September 26th at 5:00pm CET / 8:00am PT. Join me for a talk about AI optimization for the Arm Cortex-M85 with Helium vector extensions. I’ll show how we run our people detection AI at a stunning 13 FPS on a microcontroller, a 3.7x speed boost vs. Cortex-M7. We will dive into Helium MVE, compare it to the traditional Cortex-M instruction set, and show Helium code for 8-bit integer matrix-multiplications, the core of deep learning models. There will be a demo on a Renesas board with Arm Cortex-M85, a preview of our AI on Arm’s Ethos-U accelerator, and more Helium-accelerated AI apps by Plumerai. Don't miss it! #arm, #peopledetection, #tinyml, #microcontrollers cc: Tobias McBride Sign up now: https://lnkd.in/eAChpPXU Promo video below:

🚀 AI and LLM Performance Soars on Rockchip RK3588 and Mixtile Blade 3 Combination In a recent showcase of technology, the Rockchip RK3588's abilities were put to the test in artificial intelligence (AI) and large language models (LLM). Utilizing the Mixtile Blade 3 SBC, which boasts 32GB of RAM, researchers explored a range of AI tasks, from object detection with Yolo v5 to LLM exercises featuring models like RedPajama-INCITE-Chat-3B-v1-q4f16_1 and Llama series chat models. The results indicated that the Mixtile Blade 3 was adept at handling YoloV5 object detection in real-time, revealing the power of the device's Arm Mali-610 GPU when running complex LLM models. RedPajama-INCITE-Chat-3B-v1-q4f16_1 stood out, displaying commendable speed and accuracy, while even the slower Llama-2-13b-chat-hf-q4f16_1 model delivered precision. These outcomes suggest that ARM architecture, when paired with capable hardware like the Mixtile Blade 3, could serve as a significant platform for future AI and LLM venture. Sources: https://lnkd.in/gZDCq2Md #AI #LLM #Rockchip #MixtileBlade3 #SBC #TechnologyUpdate #MachineLearning #ArtificialIntelligence #GPU #ARM

Super recommended simple guide for people interested in learning Generative AI by doing experiments in their own PC! 🤖 In this case I played with image generation using Stable Diffusion to make the generation faster as explained by the guys of Nvidia here: https://lnkd.in/dUyFeFGC Using models published in HuggingFace.co This small experiment showed me the following result using only the hardware of my laptop 😎 😲

Outstanding visualisation on tokens, parameters, layers and transformers. How Gen AI and LLMs moved to become matrix weight calculations and how it moved the computing needs to GPUs. Few of the questions answered visually. What is the tally on GPT 175 billion - word embedding, unemedding and softmax.. pretty fascinating. Around 30 mts. Every second worth it. https://lnkd.in/g5nJYwsa

The new Blackwell GPU is here to tackle the AI industry’s challenges. With its advanced NVLink and Resilience Technologies and the new Tensor Cores and the TensorRT-LLM Compiler are revolutionizing the AI landscape by slashing LLM inference operating costs and energy consumption by an astonishing 25 times. Will this efficiency gain is pivotal in addressing the current GPU shortage? Time will tell soon. #AI #BlackwellGPU #Innovation