SanthoshKumar R的动态

OmniVision-968M: World's Smallest Vision Language Model OmniVision is a compact, sub-billion (968M) multimodal model for processing both visual and text inputs, optimized for edge devices. Improved on LLaVA's architecture, it features: - 9x Tokens Reduction: Reduces image tokens from 729 to 81, cutting latency and computational cost. - Enhanced Accuracy: Reduces hallucinations using DPO training from trustworthy data. https://lnkd.in/g44cqAmr

Redesigned the Hell Fire Array IP's alignment components to reduce data movement within the DDS (Data Delivery Subsystem) to improve performance of the IP under IS, WS and OS Dataflow modes. Now the alignment components have the capacity to understand the flow and lock in data without DDS needing to reload data from internal buffers / computation cycle. Hell Fire SoC Project - https://lnkd.in/grAZT5MF #SoC #hardware #vlsi #machinelearning #ai #accelerators #design #rtl #microarchitecture

Good to know

YOLOv10 object detection model just got released, and while unofficial releases are generally not as interesting as official ones, the End-to-End optimization of the this work with the removal of the NMS caught my eye as an optimization not only in compute, but also in deployment overhead/complexity reduction. I wrote up a small medium blog on it for anyone wanting to understand this concept: https://lnkd.in/dwj_smGE

Q: Why Precoders "A mathematical view". Matrix Multiplication nothing else !! Ans: suppose after Layer mapping we have data of shape (4,1024) Here, 4 is the number of layers and 1024 is number of samples in each layer. Also consider we have total 64 antennas at the gNB. Hence, precoder will apply matrix multiplication such that : (64,1024) = (64,4)@(4,1024) [(64,4) is size of precoder, @ is matrix multi] so, applying precoding is mathematically performing matrix multiplication in order to map layers to antenna ports. #5G #NR #OFDM #MIMO #PRECODER #LAYERMAPPING #3GPP

To verify the capabilities of Asterfusion Intelligence Open Network, let’s build a Kubernetes inference cluster that can run the Llama3.1-405B model.

Curious about 400G tech? Our Crash Course Video breaks down the essentials and demonstrates its revolutionary impact on data centers and network infrastructures. A must-watch for tech enthusiasts and pros alike. Catch it now! #TelecomTech #PrecisionOT #400GTechnology

SA, SAM2 and EfficientTAM Over the last few months, there’s been a lot of activity on foundational models for promptable segmentation of?images and videos. The Segment Anything Model (SAM)?- proposed last year - had trouble with?efficiently processing a large number of frames that arise for video data. In August, the Segment Anything Model 2?(SAM 2) was introduced giving a unified model for processing both image and video data.?While being the state-of-the-art for a wide range of segmentation tasks, the model includes a very large image encoder (~ 80M) and?an expensive memory module, making it inefficient for mobile deployment. EfficientTAM, which appeared late November, aims to address this by : 1. Replacing the hierarchical image encoder by a standard ViT image encoder. 2. Introducing an efficient memory module with faster cross-attention - one that uses the locality of spatial memory embeddings. In the attached figure, the two architectures are shown for a quick comparison — SAM 2 on top, EfficientTAM on the bottom. (Figures are reproduced from the respective papers?https://lnkd.in/eJMhvpYA and https://lnkd.in/ekiU5Dgi)

We use #ML to model turbulance?in our #FLEXI #CFD simulation of air flow over a plane wing travelling near the speed of sound ?? https://lnkd.in/ekwz5BGh Learn how our own Andrea Beck takes this a step further with #Relexi in this article from HLRS - High-Performance Computing Center Stuttgart and what future #HPC systems mean for this work. ?? https://lnkd.in/e75bsvgV

Today from Bynet Data Communications & NVIDIA forum TLV 2024 It was very interesting to hear about the new products and technologies and the improvement of the systems in the company’s