0 views collected around this technical thread.
The article reviews two recent ISCA 2025 papers—FAST and Neo—that introduce hardware and GPU‑based accelerators employing hoisting, KLSS, and Tensor Core optimizations to significantly boost the performance of fully homomorphic encryption workloads.
A former Nvidia engineer, working through the GIMPS distributed project and leveraging thousands of GPUs across dozens of data centers, confirmed that 2^136279841−1—a 41,024,320‑digit Mersenne prime—is the largest known prime ever found, surpassing the previous record by over 1.6 million digits.
This engineering‑focused document reviews transformer fundamentals, derives precise FLOP and memory formulas for attention and feed‑forward layers, defines the MFU performance metric, analyzes memory components and parallelism strategies, examines recent architecture variants such as MQA, GQA, sliding‑window attention and MoE, and provides practice problems applying these calculations.
This article provides a comprehensive overview of the core components and terminology of large‑scale GPU computing, covering GPU server architecture, PCIe interconnects, NVLink generations, NVSwitch, high‑bandwidth memory (HBM), and bandwidth unit considerations for AI and HPC workloads.
Over the past two decades, AI infrastructure has evolved from early distributed storage and MapReduce to GPU programming, modern package managers, in‑memory processing, deep‑learning frameworks, parameter servers, AI compilers, synthetic data pipelines, open‑source model hubs, and today’s large‑scale Kubernetes‑based clusters, forming the essential foundation for every breakthrough.
Chrome’s team has released the first implementation of the WebGPU API in Chrome 113 Beta, enabling high‑performance 3D graphics and data‑parallel compute on supported ChromeOS, macOS, and Windows platforms, with broader platform support and library integrations slated for later this year.
At the GTC conference, the talk presents Alibaba Cloud’s heterogeneous computing platform and introduces the Open Deep Learning API (ODLA), then details how CUTLASS‑based operator fusion dramatically accelerates attention and MLP layers in large‑scale recommendation models, achieving multi‑fold performance gains in production.
The article explains how massive parallel pixel calculations demand GPUs, whose high cost and inflexibility are solved by Tencent Cloud’s elastic, virtualized GPU services—including vGPU, qGPU, TACO abstraction, and spot instances—delivering up to 16 EFLOPS for AI, scientific, graphics, and video workloads.
As mobile CPUs plateau, developers increasingly use OpenCL to harness Android GPUs like Qualcomm Adreno and Huawei Mali for heterogeneous computing, leveraging its platform, execution, and memory models to write portable kernels—illustrated by a simple array‑addition example that demonstrates device initialization, kernel creation, buffer management, and parallel execution.
This article shares personal experiences and insights from using Tencent's TI-ONE machine learning platform in the 2020 Tencent Advertising Algorithm Competition, covering platform features, development modes, resource management, and lessons learned for future participants.
Since 2016 DiDi has evolved its machine‑learning platform from isolated, workshop‑style GPU servers to a cloud‑native, Kubernetes‑driven architecture that unifies resource management, introduces custom parameter‑server and serving frameworks, provides autotuning, external SaaS offerings such as Elastic Inference and JianShu, and aims for a 3.0 unified internal‑external AI marketplace.
iQIYI built a cloud‑based deep‑learning training platform called Jarvis, replacing the initial Runonce service, by containerizing GPU tasks, adopting Ceph S3 storage with FUSE, optimizing data pipelines, and addressing compute, storage, and networking challenges to improve scalability and reduce GPU idle time.
Fibonacci is JD.com's enterprise-grade serverless function platform combining container technology, serverless architecture, and event-driven mechanisms to provide high-availability, high-efficiency, and high-usability FaaS solutions.
The article surveys how NVIDIA's GPU ecosystem and the Intersect360 study reveal that among the 50 most common high‑performance computing applications, 34 already support GPU acceleration, highlighting the growing importance of GPU computing across scientific, engineering, and business fields.