Amy Chauvin Garcia’s Post

🏆 And the winner of the Embedded Awards 2024 goes to… Ambiq's Apollo510 MCU! Running on Arm Cortex-M55, it boasts up to 30x better power efficiency and 10x faster performance than a typical Cortex-M4 with FPU. This winning tech not only enables faster processing, it also opens the door to new local AI opportunities on billions of devices. Listen in as Bryan Chase talks us through Ambiq's exceptionally efficient Arm-based SoCs at #EmbeddedWorld. #ew24

Alhamdulillah! Thrilled to announce the successful defense of our final year project, where we delved deep into optimizing vision transformers for embedded devices! 🚀 Navigating the complexities of deploying a large VIT model with 87 million parameters onto edge devices like Raspberry Pi, Jetson Nano, and TX2 was no small feat. Yet, with the combined efforts of my stellar team - Hamza Javaid and Rakhmeen Gul - we rose to the challenge. Utilizing tools such as ONNX Runtime, TFLite, PyArmNN, and OpenVINO, we pushed the boundaries, surpassing CPU capabilities. Our journey was marked by hurdles, but together, we conquered them. This project is a testament to our dedication, teamwork, and passion for innovation in edge computing and machine learning deployment. #EdgeComputing #MachineLearning #EmbeddedDevices #FinalYearProject #Optimization #FPGATechnology #NUST"

Alhamdulillah! Excited to share that we successfully defended our final year project, where we delved deep into optimizing vision transformers for embedded devices! 🚀 Navigating the complexities of deploying a large VIT model with 87 million parameters onto edge devices like Raspberry Pi, Jetson Nano, and TX2 was no small feat. Yet, with the combined efforts of my stellar team - Talha Israr and Hamza Javaid- we rose to the challenge. Utilizing tools such as ONNX Runtime, TFLite, PyArmNN, and OpenVINO, we pushed the boundaries, surpassing CPU capabilities. Our journey was marked by hurdles, but together, we conquered them. This project is a testament to our dedication, teamwork, and passion for innovation in edge computing and machine learning deployment. #EdgeComputing #MachineLearning #EmbeddedDevices #FinalYearProject #Optimization #FPGATechnology #NUST"

3 petaFLOPS According to Hugging Face, just one of the AI systems (basically a supercomputer) that my CEO Darrick Horton 🌊 gave me to test has over 3 PetaFLOPS (3074 TFLOPS), water-cooled, of course 😏. Eight 750W AMD MI300X GPUs connected via a RoCE fabric. Later tonight, I'll be doing more bandwidth testing, but with TWO nodes. That's 16 GPUs using RDMA over Converged Ethernet at the same time for an estimated 6 PetaFLOPS (>6000 TFLOPS). I'll post benchmarking numbers once I'm allowed too 😁 ...until then, you'll just have to imagine how mind-blowingly fast TensorWave hardware is. Or you can test it for yourself at https://tensorwave.com/ Try the HuggingFace Calculator at https://lnkd.in/g79Y44GQ

#Amicro #AI #Chip Roadmap outlines a progressive development in AI processing power and #camera capabilities. Starting with the AM3005, featuring a 0.5T NPU, H.264/H.265 encoding, and support for 3 cameras with 1Gb DDR3, the roadmap evolves to the AM3100N, which upgrades to a dual-core 1.5GHz #CPU, 2.0T #NPU, and FHD display support with 2Gb DDR3. The AM3200 further enhances AI performance with a 4.0T NPU and AI-ISP, supporting 4 cameras and external DDR. At the high end, the AM3300 delivers a powerful 10.0T NPU, 10-camera support, 4K display, and advanced features like HDR, stitching, and USB3.0/PCIe/SATA interfaces for next-gen AI applications.

🕵♀️ Have you read about our second newsletter retracing CENTRIC first-year project results and more❔ If not, it's time 👉 https://lnkd.in/gSEVM6mx In this 2️⃣ newsletter edition, we present selected project achievements on particular aspects of designing the novel CENTRIC AI-supported air interface, followed by a presentation of the latest project publications 🗞 and recent dissemination activities such as: - Nokia Separate Training Framework for #AIML-based CSI Compression - Adaptive Non-Uniform #Quantization for #CSI "#Compression - NVIDIA #neuralnetwork (NN)-based multiuser multiple-input multiple-output (MU-#MIMO) receiver - Safe Model Predictive Control via Reliable Time-Series Forecasting - CENTRIC Project participated at the #FUSECO Forum

If you need ultra-low processing and streaming latency, unnoticeable to the human eye, take a look at this video. With MRTech SK IFF SDK it is possible to achieve 35ms Glass-to-Glass latency by receiving FullHD raw images from a camera, processing (including H.265 encoding) on a NVIDIA Jetson Orin system and streaming to a receiving computer (also, optionally, powered by IFF SDK) using the RTSP (or alternatively WebRTC) network protocol. This kind of performance is required for real-time interactive applications, such as tele-operation, augmented reality, etc. The setup also demonstrates IFF SDK capability to operate with a wide range of cameras - various hardware interfaces (PCIe, USB, Ethernet, MIPI), vendors (Basler AG, Leopard Imaging Inc, LUCID Vision Labs, Inc., XIMEA - Imaging beyond the Standard. ) and software APIs (XIMEA API, GenTL, Video4Linux). This demo case will be running in our booth 8B13 at VISION 2024, so come and see it in action! Or get a trial IFF SDK package and try it yourself: https://mr-te.ch/try #mrtech #iffsdk #lowlatency #imageprocessing #vision24 #nvidiajetson

We at MRTech SK like to test the limits of current technology to always stay on the cutting edge. Today's challenge was to see, if we could cut the 35 ms glass-to-glass latency (our previous result) almost in half, down to 20 ms. Spoiler alert - it is achievable with the right hardware and MRTech IFF SDK! Take a look at our latest experiment with NVIDIA Jetson AGX Orin and fast machine vision camera from XIMEA - Imaging beyond the Standard. 10-bit images are acquired from the camera and put through our standard CUDA processing pipeline. FullHD H265 video is then streamed over Ethernet to the receiving laptop, also running IFF SDK. When we freeze the image in the viewer, glass-to-glass latency can easily be seen by comparing the clock timestamps. They show stable 20 ms time difference, and don't forget this is remotely-viewed stream! Ask us for more details either in person at VISION 2024 or in comments section below. #iffsdk #lowlatency #mrtech #ximea #nvidia #agxorin #machinevision #visionstr

At #GTC2024, amidst the futuristic innovations and groundbreaking technologies, one booth caught the eye of StorageReview.com: Boston Limited, one of the UK’s premier system integrators. Partnered with ZutaCore, they showcased something that seemed almost against nature—boiling a GPU to cool it! ZutaCore’s revolutionary direct-to-chip liquid cooling hinges on phase change. This technology is a game-changer in managing the increasing heat generation from today's high-powered NVIDIA GPUs. In an engaging conversation, Jordan Ranous from StorageReview.com sat down with Erez Freibach, Co-Founder and CEO of ZutaCore, exploring the specifics of this innovative cooling technique and its potential impact on the future of computing. Their discussion shed light on how direct-to-chip liquid cooling challenges conventional cooling methods and offers a sustainable and efficient solution for the high-performance demands of modern technology. #GTC24 #Innovation #SustainableTech #LiquidCooling #ZutaCore #ai #hpc #zeroemissions #sustainability #datacenters https://bit.ly/3PGjHMS

🚀 Unlock the Full Potential of Llama 3.2 on Intel Hardware! 🚀 With the release of Llama 3.2, accessing the latest AI advancements has never been easier! Thanks to the seamless integration of OpenVINO™ and Optimum Intel, you can now compress, optimize, and run Llama 3.2 on your local Intel hardware in just 3 steps: 🛠 Step 1: Download the OpenVINO GenAI Sample Code (https://lnkd.in/gjKjbhHa) 🛠 Step 2: Install the Latest Build and Dependencies 🛠 Step 3: Download and Export Llama 3.2 with NNCF Here’s what I got running the inferencing on the integrated GPU of my Core Ultra7 AI PC! And it worked so well on my ARC A770 dGPU too!