Technology

Microsoft deploys world's first 'supercomputer-scale' GB300 NVL72 Azure cluster — 4,608 GB300 GPUs linked together to form a single, unified accelerator capable of 92.1 exaFLOPS of FP4 inference

Source

Tom's Hardware

Published

Oct 10, 2025

TL;DR

AI Generated

Microsoft has introduced the world's first large-scale GB300 NVL72 supercomputing cluster on its Azure cloud platform, featuring 4,608 GB300 GPUs connected by NVLink 5 switch fabric and Nvidia’s Quantum-X800 InfiniBand networking fabric. Each rack in the cluster boasts a memory bandwidth of 130 TB/s and 800 Gb/s of interconnect bandwidth per GPU. The cluster is dedicated to OpenAI workloads, enabling faster model training and deployment for advanced reasoning models. Nvidia's Blackwell Ultra GPUs and Microsoft's Azure platform are at the forefront of this deployment, with plans for more clusters worldwide.

Read Full Article

Amazon unveils 'Resilient Network Graphs' data center network that cuts hardware by 69% and boosts throughput by 33% — now the default for most AWS workloads

Amazon has introduced a new data center network architecture called Resilient Network Graphs (RNG), boasting a 33% increase in throughput and a 40% reduction in power consumption. This design, based on random graph theory, replaces traditional hierarchical structures with a flatter, more efficient network. By using 69% fewer networking devices, Amazon anticipates up to 45% cost savings and billions in reduced infrastructure expenses across its cloud services. RNG is already in use in various data centers and is set to become the standard for future deployments, offering improved performance and sustainability for AWS workloads.

Tom's Hardware•

2 weeks ago

Microsoft launches Cloud‑Initiated Driver Recovery for remote rollback of faulty updates — no user action or OEM intervention will be needed to handle broken drivers delivered via Windows Update

Microsoft has introduced a new feature for Windows called Cloud-Initiated Driver Recovery (CIDR) that allows the remote rollback of faulty drivers to a previously known good version without user or OEM intervention. This feature aims to address issues caused by bad drivers pushed through Windows Update, which have historically led to user inconvenience and manual intervention. The recovery process is initiated from the Hardware Dev Center and executed through the Windows Update pipeline, leveraging existing infrastructure. CIDR is currently undergoing validation and testing and is expected to support the Hardware Dev Center publishing process automatically starting in September.

Tom's Hardware•

4 weeks ago

AMD to broaden and specialize EPYC CPUs, already working on Zen 7 architecture — increased customization to better address evolving AI and cloud needs

AMD is planning to expand and specialize its EPYC CPUs to better cater to the diverse needs of modern data center workloads, including AI and cloud requirements. The company is already working on the Zen 7 architecture to offer increased customization for different workload-specific segments. AMD's CEO, Lisa Su, highlighted the shift towards more segmented EPYC products, potentially featuring different core/cache/interconnect configurations tailored for specific tasks like inference clusters and low-latency AI. The company aims to provide a broad portfolio of CPUs, acknowledging the need for differentiation in general-purpose operations, head nodes, and agentic AI tasks. With the upcoming Zen 6 microarchitecture, AMD plans to introduce CPUs with up to 256 cores for general-purpose servers and specialized processors for AI infrastructure.

Tom's Hardware•

1 month ago

Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year — analyst says bear thesis is 'garbage'

Google, Microsoft, Meta, and Amazon are projected to collectively spend $725 billion on capex in 2026, a 77% increase from last year. Google saw strong cloud revenue growth, while rising memory chip prices impacted spending forecasts for Microsoft and Meta. Analysts believe the AI economy is robust, dismissing bearish perspectives. Alphabet's net income surged, with Google Cloud outpacing competitors and the company's capex guidance rising to $190 billion. Shares of Alphabet rose 7% after hours, pushing the company towards a record $4.3 trillion market valuation.

Tom's Hardware•

2 months ago

Amazon unveils 'Resilient Network Graphs' data center network that cuts hardware by 69% and boosts throughput by 33% — now the default for most AWS workloads

Tom's Hardware•

2 weeks ago

Microsoft launches Cloud‑Initiated Driver Recovery for remote rollback of faulty updates — no user action or OEM intervention will be needed to handle broken drivers delivered via Windows Update

Tom's Hardware•

4 weeks ago

AMD to broaden and specialize EPYC CPUs, already working on Zen 7 architecture — increased customization to better address evolving AI and cloud needs

Tom's Hardware•

1 month ago

Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year — analyst says bear thesis is 'garbage'

Tom's Hardware•

2 months ago

Microsoft deploys world's first 'supercomputer-scale' GB300 NVL72 Azure cluster — 4,608 GB300 GPUs linked together to form a single, unified accelerator capable of 92.1 exaFLOPS of FP4 inference

TL;DR

Similar Articles

Amazon unveils 'Resilient Network Graphs' data center network that cuts hardware by 69% and boosts throughput by 33% — now the default for most AWS workloads

Microsoft launches Cloud‑Initiated Driver Recovery for remote rollback of faulty updates — no user action or OEM intervention will be needed to handle broken drivers delivered via Windows Update

AMD to broaden and specialize EPYC CPUs, already working on Zen 7 architecture — increased customization to better address evolving AI and cloud needs

Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year — analyst says bear thesis is 'garbage'

We use cookies

Microsoft deploys world's first 'supercomputer-scale' GB300 NVL72 Azure cluster — 4,608 GB300 GPUs linked together to form a single, unified accelerator capable of 92.1 exaFLOPS of FP4 inference

TL;DR

Similar Articles

Amazon unveils 'Resilient Network Graphs' data center network that cuts hardware by 69% and boosts throughput by 33% — now the default for most AWS workloads

Microsoft launches Cloud‑Initiated Driver Recovery for remote rollback of faulty updates — no user action or OEM intervention will be needed to handle broken drivers delivered via Windows Update

AMD to broaden and specialize EPYC CPUs, already working on Zen 7 architecture — increased customization to better address evolving AI and cloud needs

Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year — analyst says bear thesis is 'garbage'