#Tensor Parallelism
138 related articles
·5 min
Nonuniform Tensor Parallelism: NVIDIA's New Fault-Tolerance Approach to Boost LLM Training Goodput
NVIDIA introduces Nonuniform Tensor Parallelism, letting GPUs bear different compute loads so training can continue without checkpoint rollback during hardware failures—boosting LLM training Goodput and fault tolerance at scale.
Read more →
·3 min
picodl: Technical Analysis of a Deep Learning Library Built from Scratch with Pure NumPy
In-depth analysis of picodl, a lightweight deep learning library built from scratch with pure NumPy. Covers forward propagation, backpropagation, gradient computation, and discusses its educational value.
Read more →
·3 min
Feasibility Analysis: Deploying DeepSeek V4 Flash on Dual DGX Spark Systems
In-depth feasibility analysis of deploying DeepSeek V4 Flash on two NVIDIA DGX Spark units offline, examining memory bandwidth, MoE communication overhead, and quantization strategies.
Read more →
·3 min
GPU Utilization at 100% Doesn't Mean Full Load: How to Truly Diagnose Training Bottlenecks
nvidia-smi showing 100% GPU utilization doesn't mean optimal training efficiency. Learn about DCGM, PyTorch Profiler, and MFU metrics for diagnosing real GPU training bottlenecks.
Read more →
·2 min
nanoAlphaZero: A Single-File JAX Implementation That Trains to 2700 Elo in 24 Hours
nanoAlphaZero is a single-file AlphaZero implementation in JAX that trains an Elo 2700+ chess model in 24 hours on a TPU v4-32. The entire RL pipeline is one JIT-compiled JAX function.
Read more →
·3 min
Cursor Agent vs Codex: Which Should Visual AI Developers Choose?
In-depth comparison of Cursor Agent Window vs OpenAI Codex for Vision AI development, analyzing large task handling, multi-file edits, debugging, and long-running tasks to help developers decide.
Read more →
·3 min
Running a Small Language Model on an $8 Chip: Pushing the ESP32-S3 to Its Limits
Exploring training and running a small language model (SLM) on an ESP32-S3 microcontroller costing just $8. Learn about model design under extreme hardware constraints, quantization strategies, and edge AI's potential.
Read more →
·3 min
Hadamard Transform: A Deep Dive into the Lightweight Deep Learning Technique That Replaces Multiplication with Addition and Subtraction
Deep dive into how the Hadamard Transform replaces matrix multiplication with only addition and subtraction for lightweight deep learning, covering FWHT principles, edge inference acceleration, and ultra-low-bit LLM quantization.
Read more →
·3 min
Weak at ML Programming? A Complete Path to Breakthrough Through Systematic Practice
A systematic coding practice path for ML practitioners who 'understand theory but can't implement,' covering math basics to deep learning components with Deep-ML platform guidance.
Read more →
·4 min
RosaicLabs and Atom RTL Reveal Hidden Shifts in the x86 Ecosystem
Deep analysis of RosaicLabs, Intel Atom core RTL licensing, and 32-Tile AMX expansion — exploring x86 architecture's open licensing and customization transformation in the AI era.
Read more →
·3 min
Training a 1.3B Parameter LLM from Scratch: Complete Workflow and Core Challenges
A deep dive into the complete workflow of training a 1.3B parameter LLM from scratch, covering Transformer architecture design, data preparation, and distributed training optimization.
Read more →
·1 min
Storage-Class Memory Revolution: GPU Memory May Leap to Multi-Terabyte Capacity
Exploring how storage-class memory technology can break through GPU memory bottlenecks, expanding single-card usable memory to multi-terabyte levels through tiered memory architecture.
Read more →
·3 min
MI355X Running Kimi K3 Benchmark: The Logic Behind Outperforming B300 in Performance per Dollar
Deep analysis of AMD MI355X running Kimi K3 with superior cost-efficiency vs NVIDIA B300, and its implications for the AI inference hardware market.
Read more →
·3 min
Dual RTX 3060 Running DeepSeek V4 Flash: IQ2_M Quantized Inference Benchmarked at 3.5 tok/s
Benchmarking DeepSeek V4 Flash on dual RTX 3060 GPUs with 96GB RAM at IQ2_M quantization achieving 3.5 tokens/sec. Covers hardware choices, 2-bit quantization techniques, and local LLM deployment optimization.
Read more →
·4 min
Multi-Purpose Home Server Build Guide: Streaming + AI Inference + App Hosting on One Machine
How to build a $500 multi-purpose home server for Jellyfin streaming, Ollama local AI inference, web app hosting, and Pi-hole ad blocking with dual RTX 3060 GPUs.
Read more →
·2 min
BackdropKit: A Privacy-First Tool for Generating Launch Assets Locally in Your Browser
BackdropKit is a privacy-first local tool for creating launch assets. It supports screenshot beautification, demo videos, and data redaction—all running in your browser with no file uploads.
Read more →
·3 min
DeepSeek-V4-Flash Local Deployment Test: $8,000 Hardware Approaches Frontier Model Performance
DeepSeek-V4-Flash-0731 scores 50 on the Intelligence Index, nearly matching the frontier model score of 51 from five months prior. We analyze local deployment, hardware requirements, and implications.
Read more →
·3 min
How Cheap Can It Be to Serve a 2.8 Trillion Parameter Model? A Deep Dive into Inference Economics
Deep analysis of the real cost of serving a 2.8 trillion parameter model. From MoE sparse activation to batching scale effects and inference optimization, revealing why model size and serving cost are less correlated than assumed.
Read more →
·3 min
You Don't Always Need a Vector Database for Vector Search — Brute Force Might Be All You Need
In-depth analysis of when brute force vector search beats vector databases. For RAG apps with under a few hundred thousand vectors, brute force offers exact recall, simpler architecture, and easier debugging.
Read more →
·5 min
The Complete Learning Path to Understanding the Kimi K3 Technical Report
A systematic learning path for understanding the Kimi K3 technical report, covering MoE, MLA, distributed training, and modern post-training techniques.
Read more →