Shuai Chen

https://lnkd.in/e6wPrBZs MediaTek is working with Nvidia to co-develop an Arm-based AI processor for personal computers (PCs). [rumoured] Design to be finalized in the 3rd quarter, with verification in the 4th quarter. The chip will be priced as high as US$300 each, and is expected to become a major part of MediaTek’s business. TSMC is expected to manufacture the chip on 3nm in the 1st half of 2025, with CoWoS packaging on the back end. MediaTek is expected to reveal details of the chip in June, possibly at Computex Taipei 2024. #mediatek #nvidia #ai #pc #Semiconductorindustry #electronics #semiconductor #technology #innovation #business #supplychain #engineering #manufacturing #computerchips #semiconductors #chips #chipmaker #foundry credit: money.udn

8

5 Comments

Current status: 🦀Writing (badly) a rust script to grab some configs. 😅Building a network of L3 multicast over multi-EVPN-VXLAN fabrics in the UK office. 🎵Listening to Caroline Polachek. 🚀All. Good. Things. 🚲Having not done much multicast for 10 years, I'm back in the world of "star-comma-gee-etc". Bit like riding a bike, this stuff. Except that bike is now riding over the top of all that EVPN-VXLAN. 💜But having to switch constantly from EVPN RTs to RPF checks & VXLAN to PIM RPTs feels like having to be some networking version of a Prince-like multi-instrumentalist . 🤡Or maybe more like a one-man band. 🤔Hmmmm. 🥁Anyway, onwards! ("Bang-bang-honk!")

29

2 Comments

Geneformer is now available as part of NVIDIA #BioNeMo framework! It's a cutting-edge #AI model trained on extensive scRNA data, using a BERT-like architecture and attention mechanisms. Useful for tasks like perturbation analysis and cell classification.

6

Subjects you need to know to succeed in an OpenVPCal Calibration: https://lnkd.in/gZbhgaTr In my recent White Balancing white paper for USC ETC, I mentioned OpenVPCal as the "complex" method for calibrating a wall to a camera. I labeled this as "complex" because if you want to calibrate your wall or display as an accurate rear projection source for your camera, I have found that it is not as user friendly as some other tools out there. However, although not as user friendly, I believe it to be the most powerful — and therefore capable — calibration tool on the market if you take the time to understand it. So, here are the critical motion picture workflow skills one needs to know to succeed: 1) Video Color Management - one must be familiar with the language and tools of video color data. It's critical to understand which color space gets chosen, when to convert it, and the tools to apply these operations throughout an entire imaging chain 2) Video Playback - one must understand how to setup monitors or LED wall displays and footage playback, especially of exr file sequences. 3) Camera Color & Exposure management - one must know some fairly advanced cinematography concepts. Ie, one must understand know how to set a proper exposure and how to apply a "clean" manufacturer camera color space to recorded footage. 4) Post Production footage management - one must know how to convert recorded camera footage into to ACES AP0 linear EXR image sequences. Each of these four steps all must work for OpenVPCal to succeed. They are, in all honesty, important concepts for anyone running a stage to understand; if you learn how to keep all these ducks in a row, you will guarantee a smooth color workflow in your rear projection setup. #virtualproduction #ICVFX #cinematography #filmmaking Thanks again to Netflix, Francesco Luigi Giardiello, & Carol Payne for this monumental effort.

32

1 Comment

Thomas Coughlin has outlined the impact of AI advancements on storage needs in his current article published by Forbes. He notes that as AI applications grow, there's an increasing demand for memory and storage solutions to manage both the extensive data used in AI training and the results utilized in inference engines. This demand benefits a variety of storage technologies including DRAM, NAND flash, HDDs, even magnetic tape (#lto), and likely, emerging non-volatile memories, particularly in endpoint AI inference applications. Quantum Corporation's #ActiveScaleColdStore provides a comprehensive storage solution tailored for AI workloads. This system integrates LTO technology with SSD, HDD into a S3 Object storage system using an innovative #tape optimized Two-Dimensional-Erasure-Code technology #2DEC , ensuring efficient handling and scalability of AI-driven data demands for low cost, high performance green computing #quantum #lto #activescalecoldstore

8

1 Comment

Convolutional Neural Networks (CNNs) are integral to Optical Character Recognition (OCR) technology, including Optical Character and Symbol Recognition (OCSR) systems. Here’s a quick look at how CNNs are used in OCSR technology: Example Workflow in OCSR Using CNNs 1. Image Input: A handwritten note is captured as an image. 2. Preprocessing: The image is normalized, binarized, and denoised. 3. Segmentation: The image is segmented into individual characters or symbols. 4. Convolutional Layers: The segmented characters are processed through multiple convolutional layers to extract features. 5. Pooling Layers: The feature maps are downsampled to reduce dimensionality. 6. Fully Connected Layers: The extracted features are flattened and passed through fully connected layers for classification. 7. Softmax Layer: The network outputs probability distributions for each character class. 8. Post-Processing: The recognized text is refined using contextual information and language models. 9. Output: The final recognized text or symbols are presented for further use or verification. By leveraging the powerful feature extraction and classification capabilities of CNNs, OCSR technology can achieve high accuracy and efficiency in recognizing and interpreting handwritten characters and symbols.

3

📢 #HighlyCited 1. GNSS-RTK Adaptively Integrated with LiDAR/IMU Odometry for Continuously Global Positioning in Urban Canyons https://lnkd.in/gnD67m24 2. Arabic Hate Speech Detection Using Deep Recurrent Neural Networks https://lnkd.in/gdPgPDBR 3. Performance of Graph and Relational Databases in Complex Queries https://lnkd.in/gcAYT5nf 4. The Vision-Based Data Reader in IoT System for Smart Factory https://lnkd.in/gHGMrdas

1

Chosen-prefix Collisions for MD5 and Applications Present a novel, automated way to find differential paths for #MD5. As an application we have shown how, at an approximate expected cost of 239 calls to the MD5 compression function, for any two chosen message prefixes P and P 0 , suffixes S and S 0 can be constructed such that the concatenated values #PkS and P 0 kS 0 collide under MD5. The practical attack potential of this construction of #chosen- #prefix collisions is of greater concern than the MD5-collisions that were published before. This is illustrated by a pair of MD5-based X.509 certificates one of which was signed by a commercial #Certification #Authority (#CA) as a legitimate website certificate, while the other one is a certificate for a rogue CA that is entirely under our control (cf. http:// https://lnkd.in/dSfdTP4d). Other examples, such as MD5- colliding executables, are presented as well. Centro de Investigación de Ciberseguridad IoT - IIoT

17

Smaller #genAI models are changing the game! Introducing #mistral-NeMo by #nvidia and #mistralAI with 128k token context window - a tool for businesses that can run on RTX GPUs which you may have at your home. Super inexpensive with no high cloud costs, keeps data private on your private computer! #nvidia #mistralai #genaiforenterprise #responsibleai

10

Would you believe Jensen Huang of NVIDIA has 60 direct reports? Because he believes that : "your contribution to the work should not be based on the privileged access to information!" and "when you give everybody equal access to information it empowers people." Imagine removing 7 layers in the hierarchy of an organization: improving communication and decision-making, and allowing for more direct interaction between the CEO and a broader group of leaders. This ensures crucial information is not lost or distorted through multiple layers of management. I stumbled on this hour-long keynote that Patrick Collison of Stripe hosted and learned some unconventional yet powerful leadership insights: 𝗗𝗶𝘀𝗰𝗼𝘂𝗿𝗮𝗴𝗲𝗺𝗲𝗻𝘁 𝗼𝗳 𝗢𝗻𝗲-𝗼𝗻-𝗢𝗻𝗲 𝗠𝗲𝗲𝘁𝗶𝗻𝗴𝘀: Promotes transparency and collective problem-solving. 𝗚𝗿𝗼𝘂𝗽 𝗙𝗲𝗲𝗱𝗯𝗮𝗰𝗸: Fosters a culture of continuous improvement and shared knowledge, leveraging collective wisdom to solve problems more effectively. 𝗜𝗻𝘃𝗲𝘀𝘁𝗶𝗻𝗴 𝗶𝗻 𝗘𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀: Jensen focuses on helping employees reach their potential rather than dismissing them for their shortcomings. 𝗧𝗼𝗿𝘁𝘂𝗿𝗶𝗻𝗴 𝗘𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀 𝗶𝗻𝘁𝗼 𝗚𝗿𝗲𝗮𝘁𝗻𝗲𝘀𝘀: Setting high standards and continuously challenging employees to exceed them. True greatness comes from enduring challenges and learning from them. His leadership style is characterized by a commitment to transparency, collective learning, and continuous improvement, creating a high-performing culture at Nvidia. Check out the link to the video in the first comment—it’s well worth the watch this weekend! Hop in and listen in High #wisdompersentence Being conventional is not the best practice. PS Here is one quote that left me wondering "People misunderstand that the best jobs are the ones that bring happiness all the time. You have to suffer, struggle, endeavour, do those hard things and work through it to appreciate what you have done!" Let me know what resonated best with you in the comments below. #Leadership #Innovation #Management #TeamSuccess

18

5 Comments

In this video from our whiteboard series CPU architect Ben Fletcher explains why performance modelling has become so crucial to the design process of leading edge CPU cores.

48

1 Comment

Got Llama 70b running at semi-playable speeds on *mostly* CPU with an NVIDIA GeForce 4060 Ti running 7 GPU layers to keep *most* active layers on-device whilst the rest remains loaded for the full-size model. Currently generating Geometry and JavaScript for me. Here's a JavaScript example:

2

Can a 7B Parameter Large Model Run on 24GB of Memory? via #TowardsAI → https://bit.ly/4b3Ouf5

5

Alastair captures one of the biggest problems with the current RISC-V spec. When I looked into this (formal spec and compliance) a while back for personal interest I was (almost) shocked at the lack of rigor and state of the tooling. I say almost because I expected it, but was still sad. What saddens me more is that this situation extends to implementation validation. The status quo is sufficiently non-rigorous that if you were building an implementation you would have to do most of the architecture validation tooling lifting yourself. This will result in a few very rigorous designs (from folks who get the problem), and at least some that are not (folks who make a mess). I expect as a result an escape in a large scale design that hits the market and breaks compatibility. This stuff really matters. Read Alastair’s blog here, then remind yourself there’s a reason why some specs are thousands of pages long:

50

4 Comments