Neysa 在孟买部署 AMD GPU 训练与推理集群

这个 8,192-GPU 集群现已投入运营,将支持大语言模型训练、微调、生成式 AI 和高吞吐推理工作负载。
该部署整合了 AMD Instinct MI350X GPU、AMD EPYC 处理器以及 AMD ROCm 软件生态。
Neysa 表示,该集群一次性完成上线,全球客户已签约锁定用于在印度开展 AI 部署的算力容量。
该基础设施旨在让客户独享加速器访问权限,并可自主掌控网络架构、存储、部署模式和服务级别要求。
客户可在裸金属部署、裸金属加 Slurm,或裸金属加 Kubernetes 之间选择。
该部署也扩展了 Neysa 的芯片无关基础设施战略。
通过 Tytan,该公司允许客户根据自身工作负载选择加速器架构和运行配置,而集群部署与运营则由 Neysa 负责。
Neysa 表示,MI350X 集群旨在同时服务印度本土企业,以及在印度建立算力能力的全球 AI 公司。
该公司还运营着 Velocis,这是一个托管式 AI 云平台,将 GPU 算力、编排、推理、可观测性和成本管理整合在一起。
"对于大规模 AI 部署而言,问题不再仅仅是有多少 GPU 可用,而在于基础设施能否规模化部署、能否稳定可预期地运行,以及能否针对具体工作负载进行配置,"Neysa 联合创始人兼 CEO Sharad Sanghi 表示。"Neysa Tytan 让 AI 算力采购方掌握主动权:专属裸金属算力容量,并可自由选择符合自身需求的芯片、架构和运营模式。随着 AMD Instinct MI350X 在 Tytan 上落地,我们正在扩大这一选择范围,同时提供全球客户在印度部署 AI 基础设施所需的规模。"
10 月,Neysa 与 ST Telemedia Global Data Centres(STT GDC)达成合作,在 STT Noida 2 设施部署可支持多达 9,000 块 GPU 的 GPU 基础设施。第一阶段预计于 2027 年第一季度上线,将通过 Neysa 的 Velocis 平台支持 AI 训练和推理。
这一多年期安排旨在服务印度企业和全球算力客户,其既定目标包括数据驻留和可预期的延迟。
2 月,Blackstone 及其共同投资方签署最终协议,向 Neysa 提供最高 6 亿美元的股权融资,该公司还计划再筹集 6 亿美元债务融资。这笔资金旨在支持其扩张计划,包括在印度各地部署超过 20,000 块 GPU。
Neysa 的这则公告也标志着 AMD 在印度扩大其 GPU 基础设施方面又迈出重要一步。
10月,Vultr宣布计划在钦奈一个20 MW的AI超级集群中部署10,000块AMD Instinct MI355X GPU。该设施预计将成为印度最大的AMD Instinct超级集群,也是全球第二大。Vultr还概述了于2027年使用AMD下一代MI455X GPU和Helios机架级架构扩展该基础设施的计划。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Neysa Deploys AMD GPU Cluster in Mumbai for AI Training and Inference)
The 8,192-GPU cluster is now operational and will support large language model training, fine-tuning, generative AI and high-throughput inference workloads.
The deployment brings together AMD Instinct MI350X GPUs, AMD EPYC processors and the AMD ROCm software ecosystem.
Neysa said the cluster was brought online in a single phase, with global customers already having contracted capacity for AI deployments in India.
The infrastructure is designed to give customers dedicated access to accelerators, along with control over network architecture, storage, deployment models, and service-level requirements.
Customers can choose between bare-metal deployments, bare metal with Slurm, or bare metal with Kubernetes.
The deployment also expands Neysa's silicon-agnostic infrastructure strategy.
Through Tytan, the company allows customers to select accelerator architectures and operating configurations based on their workloads, while Neysa manages cluster deployment and operations.
Neysa said the MI350X cluster is intended to serve both Indian enterprises and global AI companies establishing compute capacity in the country.
The company also operates Velocis, a managed AI cloud platform that combines GPU compute, orchestration, inference, observability and cost management.
"The question for large AI deployments is no longer simply how many GPUs are available. It is whether the infrastructure can be deployed at scale, operated with certainty and configured for the workload," said Sharad Sanghi, Co-Founder and CEO of Neysa. “Neysa Tytan gives AI offtakers control: dedicated bare-metal capacity and the freedom to choose the silicon, architecture and operating model that fit their requirements. With AMD Instinct MI350X on Tytan, we are expanding that choice while delivering the scale global customers need to deploy AI infrastructure in India.”
In October, Neysa partnered with ST Telemedia Global Data Centres (STT GDC) to deploy GPU infrastructure supporting up to 9,000 GPUs at the STT Noida 2 facility. The first phase is expected to go live in Q1 2027 and will support AI training and inference through Neysa's Velocis platform.
The multi-year arrangement is designed to serve Indian enterprises and global compute customers, with data residency and predictable latency among its stated objectives.
In February, Blackstone and co-investors entered into definitive agreements to provide up to $600 million in equity financing to Neysa, with the company planning to raise another $600 million in debt. The funding was intended to support its expansion plans, including deploying more than 20,000 GPUs across India.
Neysa's announcement also marks another major step for AMD in scaling its GPU infrastructure in India.
In October, Vultr announced plans to deploy 10,000 AMD Instinct MI355X GPUs in a 20 MW AI supercluster in Chennai. The facility is expected to become India's largest AMD Instinct supercluster and the second-largest globally. Vultr has also outlined plans to expand the infrastructure in 2027 with AMD's next-generation MI455X GPUs and Helios rack-scale architecture.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 Analytics India Magazine(Neysa Deploys AMD GPU Cluster in Mumbai for AI Training and Inference),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。