#GPU
25 related articles
DeepSeek Open-Sources Inference Engine: The Inside Story of Tripling GPU Value
DeepSeek open-sources DeepInfer, its production inference engine, using pre-fill/decoding separation, wide expert parallelism, and 96% KV cache hit rate to lift revenue per GPU-hour from $12.91 to $38 with 84%-92% gross margin.
How Kubernetes Becomes Day-One Ready for NVIDIA Vera Rubin
NVIDIA's technical blog explains how to make Kubernetes AI-Ready for the new Vera Rubin GPU on day one by converging driver, kernel, runtime, and network adaptation into a compute-domain abstraction, reducing manual configuration and error risk.
Wireless GPU Hands-On: A Laptop Calls an RTX 5090 Over the Air
A YouTube video shows a wireless GPU that lets a laptop call an RTX 5090 over Wi-Fi—no cables needed—running Cyberpunk, with claims of a 410TB memory pool and Wi-Sci architecture.
Fiber Lasts Decades, GPUs Only Three Years: The AI Boom Is Not the Dot-Com Bubble
Hacker News discussion argues the AI boom differs from the dot-com bubble: infrastructure like fiber lasts decades, while GPUs have an economic life of only about three years—an asset structure difference that defines distinct risk profiles and investment logic for the two waves.
Why Does AI Compute Favor Rectangles? Modal CEO Breaks Down Agent Cloud Economics
Modal CEO Erik Bernhardsson explains why AI compute always comes in rectangular form and discusses elastic GPU scheduling, sandbox cold start optimization, and agent cloud economics, revealing the real trade-off between utilization and cost structure.
Samsung's HBM Capacity Sold Out: Is a GPU Price Surge Coming?
Samsung's 2027 HBM capacity is sold out, with the market expecting HBM prices to rise 200% and DRAM to see double-digit increases. This article analyzes how memory price hikes pass through to GPUs and consumer hardware, and the real impact on gamers and developers.
Scaleway's Dual-Country GPU Zones Go Live: AMD and NVIDIA Running in Parallel
Scaleway announces Stockholm AMD MI355X and Copenhagen NVIDIA B300 clusters, Milan reaching three availability zones, and a 40+ model catalog. We break down the dual-chip strategy's impact on buyers and what can be priced now.
Why Your Computer Can't Run AI Models: Video Memory Is the Real Barrier
The 8B in AI models means 8 billion parameters, not 8GB. This article explains the difference between model parameters and storage size, how quantization compresses size, and why running AI models requires GPU video memory rather than a hard drive or system memory.
NVIDIA RTX 50 Series GPU Prices Soar in Japan: Up to 41% Higher Than US
Tokyo retail prices show NVIDIA RTX 50 series GPUs selling in Japan at 5.66% to 41.1% above Amazon US, with RTX 5060 Ti carrying the largest premium. Memory chip shortages and AI compute demand squeeze consumer capacity, and RTX 5090 is out of stock in Japan.
RTX 5090 Scalped to $5,000? AI Companies Are Buying Gaming GPUs by the Pallet
Why is the RTX 5090 scalped to $5,000? AI companies are buying gaming GPUs by the pallet, causing a consumer GPU shortage and price surge.
Rumor: NVIDIA May Discontinue RTX 5090, Redirecting GB202 Chips to RTX PRO Professional Cards
Rumors suggest NVIDIA may discontinue the RTX 5090, redirecting GB202 chip supply to higher-margin RTX PRO professional cards. Analysis of the business logic and potential impact on gamers and AI developers.
CUTLASS Python: NVIDIA's New Tool for Accelerating GPU Kernel Development
NVIDIA engineer introduces CUTLASS for Python updates: new Qt extensions, CUTLASS primitives, compiler diagnostics, and static checks for high-velocity GPU kernel development for both developers and AI agents.
AICR v1.0 Released: An Open, Verifiable GPU Cluster Configuration Reference
NVIDIA's AICR v1.0 delivers an open, stable, and verifiable GPU cluster configuration reference for Kubernetes, solving version compatibility challenges across kernels, drivers, CUDA, and more.
NVIDIA Green Contexts: Fine-Grained Control Over GPU Resource Allocation
NVIDIA Green Contexts lets developers programmatically partition GPU SM resources within a single process, isolating latency-sensitive and throughput-intensive workloads for predictable performance.
Why Do AI Data Centers Consume So Much Power? From GPU Heat to the Nuclear Energy Dilemma
Why do AI data centers consume so much power? This article unpacks server room operations, GPU cooling challenges, integrated photonics solutions, and why U.S. data center energy use could equal eight New York Cities by 2028.
TileLang: A DSL for Writing High-Performance GPU Kernels in Python
TileLang is a Python-syntax DSL for high-performance GPU/CPU/accelerator kernel development. Learn about its tile abstraction, cross-hardware design, and use in LLM operator optimization.
Why Real-Time AI Inference Costs 56x More Than It Should — And How Persistent Kernels Fix It
Why does serving real-time AI cost 56x more than it should? We break down continuous vs. request-based inference, and how persistent GPU kernels, synchronous batching, and compile-time scheduling fix it.
Building a 64GB Local AI Inference Server with 4x Tesla T4 GPUs: A Full Breakdown
A Reddit user details building a local AI inference server with 4x Tesla T4 GPUs (64GB VRAM), covering hardware selection, llama.cpp+OpenWebUI stack, and why CPU core count isn't the bottleneck.
Buy or Rent H200 GPUs? A Real Break-Even Analysis
Buy or rent an 8-GPU H200 server? A real break-even model: $370K purchase vs $4.40/GPU-hour rental median, payback at 100/60/40% utilization, plus 4 hidden costs.
A Winter GPU Joke: Can Your Graphics Card Really Be 100% Efficient?
A viral Reddit joke about GPUs being "100% efficient" in fall highlights the real physics of GPU heat output, waste heat recovery in data centers, and the true meaning of performance per watt.