Global GPU Computing News【20251018】

Microsoft Deploys World's First GB300 Supercomputing Cluster: Microsoft Azure has officially deployed the NVIDIA GB300 NVL72 supercomputing cluster, built specifically for OpenAI. This cluster integrates over 4,600 Blackwell architecture GPUs, delivering peak FP4 tensor core performance of up to 1440 PFlops. It is designed to reduce the training time for trillion-parameter models from months to weeks.

OpenAI and AMD Enter Chip Cooperation Agreement: OpenAI has signed a multi-year AI chip supply agreement with AMD. Starting in the second half of 2026, AMD will gradually deliver hundreds of thousands of MI450 series GPUs built on a 2-nanometer process, representing a total deployment scale of approximately 6GW of computing power.

Oracle Unveils AI Supercomputer with 16 ZettaFLOPS Peak Performance: Oracle launched its next-generation AI supercomputer, "OCI Zettascale10", claiming a peak performance of 16 ZettaFLOPS. The system is designed to connect hundreds of thousands of NVIDIA GPUs across different data centers to form multi-gigawatt-scale clusters.

NVIDIA Launches DGX Spark Desktop AI Supercomputer: NVIDIA announced that its DGX Spark desktop AI supercomputer became available on October 15. The device is based on the GB10 Grace Blackwell superchip, delivers 1 petaflop of FP4 sparse AI computing power, and supports local inference for 200-billion-parameter models.

NVIDIA Showcases Next-Generation Kyber Rack Platform: NVIDIA showcased its Kyber rack-level technology platform. This platform adopts an innovative vertical blade design, plans to support the integration of up to 576 Rubin Ultra GPUs per rack by 2027, and will utilize 800V DC power to improve energy efficiency.

Alibaba Cloud's Domestic Heterogeneous GPU Platform Wins "Computing China" Award: The "Key Technologies and System for Domestic Heterogeneous GPU Cloud Platform", jointly developed by Alibaba Cloud and Zhejiang University, was honored with the "2025 Computing China · Annual Major Achievement" award. The platform has successfully built a 10,000-card hybrid heterogeneous computing cloud platform.

Shanghai Issues Action Plan to Strengthen Edge AI Chip Layout: Shanghai issued the "Shanghai Smart Terminal Industry High-Quality Development Action Plan (2026-2027)", proposing to actively develop edge-side GPU chips based on server computing chip capabilities.

Peking University Team Successfully Develops High-Efficiency Analog Matrix Computing Chip: Peking University successfully developed a high-precision, scalable analog matrix computing chip based on resistive random-access memory (RRAM). When solving specific scientific problems, its computational throughput and energy efficiency are improved by hundreds to thousands of times compared to current top-tier GPUs.

Moore Threads Passes STAR Market IPO Review: Domestic full-featured GPU maker Moore Threads passed the STAR Market listing committee review for its IPO. It plans to raise approximately 8 billion yuan to fund the development of its new generation of independent and controllable AI training-and-inference chips.

Alibaba Cloud Releases Panjiu 128 Ultra Node AI Server: Alibaba Cloud released the Panjiu 128 Ultra Node AI server. A single cabinet supports 128 AI computing chips, provides petabit-per-second-level bandwidth, and can improve inference performance by approximately 50% at equivalent computing power.